draftwright 0.4.35__tar.gz → 0.5.0rc2__tar.gz

draftwright-0.5.0rc2/CHANGELOG.md

@@ -0,0 +1,3198 @@
+# Changelog
+
+## Unreleased
+
+### Added
+
+- A hash-pinned AP242 CTC-01 canary now proves the full Pareto-first layout loop against
+  a reviewed 80-measurement denominator: read typed finding remedies, apply only two public
+  declaration-side overrides, replay once, and establish legibility-only dominance while 12
+  unresolved claims remain carried and certification remains unavailable (#1753). Detailed
+  diagnostic prose remains optional presentation derived from structured evidence, never a
+  comparison identity or decision input.
+- Declared slot width and length dimensions now expose bounded, feature-relative lanes through
+  `DimensionIntent.place(lane=...)` and declaration-scoped
+  `layout_override(..., parameter=..., lane=...)`. Generated scripts and report-v8 evidence
+  retain the layout-only policy; the shared solver admits only wholly interior or exterior,
+  page-safe, ink-clear candidates and reports bounded blocker evidence when a requested lane is
+  infeasible. Automatic drawings and scripts without lane intent remain unchanged (#1757).
+- Declared sheets now support append-only, build-scoped `layout_override(..., side=...)`
+  edits plus structured capability and validation queries. Overrides select only a semantic
+  corridor, retain declaration identity through generated-script replay, and appear as
+  layout-only resolution evidence in declared report v8; coordinates and engineering meaning
+  remain unchanged, and sheets without overrides preserve existing behaviour (#1757).
+- Raw automatic report assessments now include a typed remediation projection that separates
+  contradictions, missing carriers, layout, importer/lowering, source ambiguity, and unavailable
+  evidence. It retains affected evidence identities, names preservation invariants, advertises
+  only bounded public actions, and leaves unknown findings unclassified without guessing (#1759).
+- Raw automatic reports now lead agents with a vector-first `lint.assessment`: per-axis status,
+  affected counts, available denominators, unavailable reasons, and stable structured reasons
+  for every `needs-attention` verdict. Its human summary is rendered from the same reason rows,
+  while legacy lint scores remain compatibility-only diagnostics and cannot select status
+  (#1609, #1618).
+- Replay comparison now matches unresolved measurement evidence by its serialized annotation,
+  table cell, or ownerless parameter identity and classifies it as carried, resolved,
+  introduced, changed, or ambiguous. Carried uncertainty keeps certification unavailable but
+  no longer erases an otherwise Pareto-dominant layout improvement; new, changed, or duplicate
+  identities fail closed (#1753).
+- Assessment comparison v2 reports independently evidenced requirement, completeness,
+  fidelity, legibility, and restraint axes plus their Pareto relation. Dominance,
+  equivalence, domination, incomparability, and unavailable evidence are derived from concrete
+  transitions rather than any legacy diagnostic, component, or audited-completeness score
+  (#1609). The v1 `decision` remains temporarily as a deprecated compatibility policy result.
+- Stadium through-slots now retain Quiddity 0.3.2's proved end radius through detection,
+  declared `Sheet` scripts, compiled measurement identity, placement, and completeness. A
+  normal-to-arc `2× R…` leader distinguishes an obround from an equal-sized rectangular slot
+  without introducing coordinate placement (#1752).
+- The Quiddity 0.3.2 adoption also retains corrected overall extents for rounded-corner
+  rectangular slots. Hole-pattern records whose members differ only along the drilling axis
+  are explicitly refused at the drafting boundary, preserving their ordinary grouped-hole
+  requirements instead of emitting a coincident, undrawable pitch pattern.
+- Feature-leader placement now has one coordinate-free drawing policy across
+  `build_drawing`, `make_drawing`, `Sheet`, generated scripts, and the CLI. The default
+  `auto` policy keeps the shared interior/exterior solve, `interior` requires proved
+  interior candidates for eligible unconstrained families, and `exterior` restores the
+  historical exterior-only layout for existing diagrams.
+- The standard title block now states drawing `UNITS` (`mm`) and the effective sheet `FORMAT`
+  (for example `A2`). Dimension-specified A-series pages receive the standard designation;
+  non-standard pages state their actual `WIDTHxHEIGHT`. Its rows share a four-line column grid,
+  avoiding visually offset dividers between adjacent fields.
+- A bounded real-part canary now proves the complete agent improvement loop on NIST CTC-01
+  (#1716): inspection-to-declaration joins, traced baseline, one sanctioned `Sheet` layout edit,
+  replay assessment, fixed-denominator comparison, and fail-closed deletion/owner/provenance
+  checks. It performs exactly two CAD builds in the existing single PR canary lane.
+- Replay assessments v2 now retain exact confirmed compiled measurement meanings, and
+  `compare_assessments()` evaluates two same-source replays as separate semantic, lint,
+  completeness, fidelity, layout, and availability deltas (#1712). Caller-fixed requirements
+  expose shared omissions; authorised design changes remain separate; no composite score is
+  produced.
+- Generated scripts now atomically write a versioned post-build replay assessment beside their
+  outputs (#1715). It binds the exact STEP and edited Python bytes, producer/run options, output
+  hashes, declaration/representation map, layout evidence, and lint from the same finalized
+  drawing. Failed replays clear only stale Draftwright-owned assessments; `--no-report` opts out.
+- Declared drawing report schema v8 publishes page/view geometry, named annotation ink and
+  public line segments, raw-lint joins, semantic remedy choices, and explicit opt-in solver
+  evidence (#1711). Coordinates are review evidence only; editable links remain final-IR
+  declarations, and an untraced or partially traced build says so instead of appearing clear.
+- STEP inspection schema v4 projects a deterministic report-local source-face roster and
+  Quiddity 0.3.1's closed graph rooted at rejected candidates (#1726). Accepted occurrences and
+  candidates carry exact defining/constituent face links; rejected roots and direct relationships
+  reconcile with aggregate lifecycle counts. The report still states that this is diagnostic
+  evidence, not precision, recall, missed-feature truth, or persistent topology identity.
+- Circular-seat recognition, declared features, and generated-script replay (#1613).
+  Diameter, open run, arc sweep, and three datum-to-axis coordinates retain each physical
+  seat's measurement identities, including counted coaxial callouts and shared axis marks.
+- Opt-in independent sheet margins and exact title-block width on `Sheet`, `build_drawing`,
+  `make_drawing`, and the CLI, including generated-script replay (#1612, #1589). Sergio's
+  25/10/10/10 mm margins and 175 mm block are explicit options; existing defaults remain.
+
+### Changed
+
+- The agent usage guide moved from `AGENTS.md` to `docs/using-draftwright.md`; `AGENTS.md`
+  now holds the contributor instructions formerly in `CLAUDE.md` (#1938).
+
+### Fixed
+
+- Assessment comparison now rejects an apparent resolution of measurement uncertainty when the
+  declaration carrying that unresolved annotation was removed. Deleting crossed GD&T ink can no
+  longer masquerade as evidence improvement; genuine resolution with the carrier retained still
+  compares normally (#1753).
+- Imported upper/lower limit dimensions now retain their typed absolute interval through the
+  compiled evidence boundary. Both bounds together satisfy the owned nominal claim, while a
+  changed/missing bound or two coincidental numbers around the nominal still produces
+  `claimed_value_absent`; count and through requirements remain separate (#1759).
+- Adopt Quiddity 0.3.1 and require its public rejected-candidate reference in the
+  evidence contract (#1726). The provider now suppresses rectangular recess candidates
+  that are not material-empty and restores passage discovery used during rotational
+  reconciliation; Draftwright's recognition and real-part regression corpora retain
+  their reviewed expectations.
+
+- Direct and generated builds use the same final IR step ladder for page layout (#1592).
+  Raw face levels owned by another feature no longer reserve a phantom step corridor at
+  the title-block clearance threshold.
+
+- Location intents now accept `.format(decimals=...)` for all selected directions (#1610),
+  including plan locations, side-drilled holes and slots. Compiler-owned labels reach
+  placement footprints, live location edits and generated scripts. Coincident marks with
+  incompatible explicit display text are refused instead of overriding an intent.
+
+- Explicit dimension precision, including `.format(decimals=0)`, now reaches geometric
+  label critique (#1620). Correct rounding is reported as information; labels outside
+  the requested rounding interval still produce `label_vs_measured`. Height and step
+  ladders retain the same compiler policy as ordinary dimensions.
+
+- **`label_vs_measured` no longer reports the drawing's own rounding as a topology error
+  (#1600).** A sheet at one decimal place prints 4.450 as `4.5`; the check then compared
+  `4.5` against 4.450 and said "possible axis swap or wrong endpoint". The comparison was
+  *relative*, so the same 0.05 mm of rounding passed silently on a 100 mm dimension and was
+  an error on a 4 mm one — it fired on smallness, not on wrongness.
+
+  A label now states its own precision: `4.5` shows one place, `4.45` two, `12` none. A
+  difference within half of the last displayed place is the label doing its job; anything
+  beyond it is still reported, and an axis swap or wrong endpoint misses by far more.
+
+  This is a prerequisite for #1602's proposed CI gate: `label_vs_measured` sits in the
+  register of codes meaning *the sheet says something untrue*, and a code that fires on
+  correct rounding cannot be hard-failed on.
+
+### Added
+
+- **`nominal_rounded` says where the sheet's precision loses the model value.** Silencing
+  the false alarm above left nothing reporting it at all, and `4.5` for 4.450 is still a
+  real 0.05 mm between drawing and model — on a part carrying 0.05 mm clearances, the whole
+  clearance. One `info` per sheet, naming the worst case:
+
+  > 7 dimension(s) print a nominal the model does not have, rounded to the sheet's
+  > precision; the largest is '15.7' for 15.6500 (0.0500 mm). Within a general tolerance
+  > this is ordinary; where a fit depends on it, raise that dimension's places with
+  > `.format(decimals=...)`
+
+  `.format(decimals=…)` cannot yet reach a **location** dimension (#1610), and on the
+  fixture this was measured against six of the seven offenders are locations — so the
+  message says so rather than prescribing a remedy that raises `ValueError`.
+
+  It does **not** guess, and does not change any rounding. `4.450` locating a hinge axis is
+  design intent; `71.595` as an overall envelope is parametric fallout no drafter would
+  print in full — and telling them apart needs to know what the part is *for*, which a STEP
+  file does not carry (the same gap as #1597). So the engine reports and the author decides:
+  `.format(decimals=…)` already sets places per dimension and survives the `--script` round
+  trip.
+
+  Measured: 7 of 25 numeric dimension labels on `whistle_frame_reference.step`; silent on
+  every golden fixture, all of which are built from round numbers.
+
+### Fixed
+- Adopt published Quiddity 0.2.9 (#1619). The frame document now retains six separate
+  hinge bores instead of a single path spanning air gaps. Three cylindrical seat occurrences
+  are reported explicitly; their drawing grammar remains pending #1613.
+- Refused hole patterns retain their members in the hole-completeness ledger (#1607).
+  The adapter records the refusal against the exact recognition run, so a six-hole group
+  is no longer reported as an unverifiable four-hole pattern and two loose holes. Declared
+  pattern mismatches still undergo the independent correspondence checks.
+
+- **A required dimension with nowhere to go now re-plans the sheet instead of being
+  reported and left (#1590).** The automatic recovery ladder — larger scales on the
+  selected page, then dropping the optional pictorial, then a larger page — already
+  existed and already refused any candidate that had not fixed the problem. It ran on
+  two symptoms: a turned part missing an axial station, and an **authored** intent that
+  failed to place. A required *envelope* dimension that found no room was neither, so
+  these drawings never entered the ladder at all — `scale_decision["attempts"]` came
+  back empty and an `overall_dim_withheld` error was reported on the first sheet tried.
+
+  `overall_dim_withheld` is deliberately not a scale blocker, and that is right:
+  reporting it through `placement_unsatisfiable` once made `build_drawing(part,
+  scale=...)` raise on parts that had always built. But that argument is about refusing
+  the scale a **caller asked for**. Choosing one automatically is a different question,
+  and one predicate was answering both.
+
+  What it changes, measured:
+
+  | part | before | after |
+  |---|---|---|
+  | stepped `Box(80, 60, 30)` | A4 1:1, overall depth withheld (error) | A4 1:1, **clean**, pictorial dropped |
+  | 80 × 60 plate, two close bores | A4 1:1, `incomplete`: unlocated hole, dropped location, withheld depth | A2 2:1, **clean**, every dimension placed |
+
+  **ADR 2 invariant 13 is amended** to match. It read "the optional isometric yields only
+  to close missing axial coverage", which had been one trigger behind the code since
+  2026-08-23 — the record was consolidated twelve days later and transcribed the narrower
+  claim. It now states the principle the triggers share: the isometric yields only to
+  preserve manufacturing completeness, never for space, crowding or preference. The list
+  itself is `builder._ISO_YIELD_TRIGGERS`, and a guard fails if it changes, so the next
+  widening is caught by a test rather than by a reader.
+
+  Dropping the optional pictorial is the first lever, which is the order #443/#1299
+  already argue for — a pictorial cannot outrank a dimension needed to manufacture the
+  part. Pin `page=` or `scale=` to opt out; a pinned sheet still reports honestly rather
+  than replanning.
+
+- **A generated `--script` reproduces a re-planned layout.** A declared build does what
+  it is told and never enters the recovery ladder, so `generate_sheet_script` pins the
+  settled page, scale and view set into the script. It used to do that only for the two
+  families whose replan could be predicted from the model; nothing predicts the new
+  trigger, because whether a mark fits is a fact about a measured sheet. The reference
+  build is now unconditional — one extra `build_drawing` per generated script, measured
+  at 0.19 s for a plain box, 0.34 s for a pocket and 1.17 s for a part that actually
+  replans — and an **undrawable** source can never be the reason a script is not
+  written, the standard the inspection sidecar beside it already meets. A deliberate
+  refusal (an unknown page size, `ScaleIncompatibilityError`, `ViewPlanIncomplete`)
+  still reaches the caller.
+
+  A script that keeps `auto_dimensions()` is not handed a pinned view set: `Sheet`
+  refuses `authored_views()` beside it, so the two together wrote a script that raised
+  when run. The settled set is reported in a comment instead.
+
+- **A bolt circle is no longer stated unless something other than the fit supports it
+  (#1596).** `EQ SP ON ø… BC` tells a machinist to work from a centre; #1595 found a real
+  part where that centre was in mid-air. Six ⌀2.4 holes in a 2×3 rectangular grid came out
+  as `4× ⌀2.4 THRU EQ SP ON ø34.4 BC` plus an unrelated `2× ⌀2.4 THRU`, the circle centred
+  on (0, −20.11) — nothing is concentric with it.
+
+  The fit was not a near miss. **Any three non-collinear points are concyclic, and so are the
+  four corners of any rectangle** — for those arrangements the residual is zero by
+  construction, so no tolerance on it can help. Nor does equal spacing rescue it: that
+  rectangle is near-square (24.30 × 24.38), so its corners sit 90.19° / 89.81° apart and
+  would pass an `EQ SP` check too.
+
+  What gives it away is the leftover callout beside it. Corroboration must now come from
+  somewhere the fit cannot reach — any one of:
+
+  1. **five or more members** — a circle through five holes could have failed to fit;
+  2. **a concentric physical feature** — a boss, spigot, central bore, or a round body whose
+     own axis is the circle's, something a machinist can indicate off;
+  3. **it accounts for every hole of its kind** — no identical hole left off the circle.
+
+  (3) is the discriminator here: six identical holes and the circle claimed four. A drawing
+  does not put two of six bolts on a different plan. It is also a question about the part
+  rather than about its angle to the axes, unlike the row/column test it replaced.
+
+  Refused patterns keep every hole: the members fall through to ordinary grouping, so the
+  #1595 part now reads `6× ⌀2.4 THRU` — which also repairs the fragmentation into 4 + 2.
+
+  **What refusing costs, measured.** `hole_requirement_outcomes` builds its pattern-member
+  set from the recogniser's patterns rather than the ones the IR adopted, so after a refusal
+  it keeps those holes out of the loose groups, finds no `PatternFeature` to join, and warns
+  `hole_requirement_unverifiable` about holes the sheet states correctly — two on this part,
+  26 issues against 23. The pre-existing oblique-pattern refusal (#971) has the same hole and
+  was simply never reachable; #1596 makes it so. Tracked as **#1607** with three attempted
+  fixes recorded, and guarded by an `xfail(strict)` here.
+
+  Refused at the recognition→IR adapter, not in the recogniser: a circle through those holes
+  is genuinely findable, and ADR 3 leaves the recogniser to report the geometry it finds.
+  Whether it may be *stated as a drafting datum* is drafting policy, which is draftwright's.
+
+### Changed
+
+- **The title block carries every ISO 7200:2004 mandatory data field.** Three had
+  no representation at all before — segment/sheet number (5.1.6), approval person
+  (5.3.4) and document type (5.3.6) — and are now `sheet=`, `approved_by=` and
+  `document_type=` on `build_drawing`, `make_drawing`, `Sheet`, the CLI and the
+  `--script` emitter. Cell widths come from the standard's own recommended
+  character counts (via `build123d-drafting-helpers>=0.15.4`), so they follow ISO
+  7200 and the font rather than a chosen proportion.
+
+  Eleven fields need four rows, so **the block is 32 mm tall rather than 16**, and
+  the band reserved for it goes 35 → 40 mm. `material` and `general_tolerance`
+  stay although neither is an ISO 7200 title-block field; they are on every real
+  drawing.
+
+- **The scale is stated outside the title block**, beside the projection symbol,
+  per ISO 7200 §4 ("presented outside the title block only when used, e.g. scale,
+  projection symbol"). It is drawn unconditionally, and the new **`scale_not_stated`
+  error** checks the rendered sheet rather than trusting the code that draws it.
+  The `SCALE` cell is gone from the block.
+
+  **Two consequences worth knowing before you upgrade:**
+
+  - A dense sheet may escalate a page size. In the golden corpus one fixture of
+    fourteen (`flange_dense`) moves A4 → A3; the other thirteen keep their page
+    and scale, and a sweep of ten ordinary shapes showed no change at all.
+  - **Dimension placement now avoids the title block (#1593)**, which it never
+    did: the block is drawn after placement, so only GD&T frames ever avoided it
+    (#481). A two-row block kept that out of reach; a four-row one does not.
+    `build_drawing(Box(40, 30, 12))` used to run its `30` dimension 2.1 mm into
+    the block with `lint()` reporting a clean sheet; a stepped
+    `Box(80, 60, 30)` put the whole `60` label inside it, reported only as an
+    `annotation_overlap` between the labels `60` and `DRAWING`.
+
+    Most displaced dimensions move to the view's opposite strip and stay on the
+    sheet. Where both strips are full the dimension is **withheld as an error
+    naming the block**, rather than drawn through it: in the golden corpus
+    `pocketed` and `prismatic_ladder` lose their overall depth at the automatic
+    A4/1:1, and place it on A3 or at 1:2. Likewise a **pinned** A4 for a
+    five-step turned shaft no longer reaches 5:1, because at that scale the front
+    view runs down past the block and the overall height cannot be placed; the
+    2:1 sheet is returned with `axial_length_missing` stated. Leaving the page
+    automatic still carries every station, on A3 at 5:1.
+
+    Choosing a page and scale without knowing which dimensions will still need
+    room is #1590.
+
+### Changed
+
+- **Exports are reproducible by default.** `reproducible` now defaults to `True`
+  on `build_drawing`, `make_drawing`, `Drawing` and `Drawing.export`, so two
+  exports of one drawing are byte-identical without asking. Pass
+  `reproducible=False` to opt out.
+
+  It was opt-in on a cost argument, and the cost turns out to be small at the
+  scale most drawings are made: on the NIST CTC-01 AP242 fixture, +0.27 s on a
+  13.4 s job (+3.2% of export, +2.0% overall). The +32% of export measured when
+  the flag was introduced was a 358-part assembly; the ordering work is one
+  bounding box and one edge walk per part, so it scales with part count and a
+  part-heavy sheet is still the case to opt out of.
+
+  Two things pushed the default over. A file that changes between runs cannot be
+  diffed, checksummed or cached — the reason to want it is not speed-shaped. And
+  neither the CLI nor `Sheet` exposes the flag at all, so every drawing made
+  through those surfaces was non-reproducible with no way to change it.
+
+  One behaviour follows from the default rather than from this change: a DXF
+  export that cannot reach the ezdxf document to pin its metadata now raises
+  rather than writing live data, because every export now claims
+  reproducibility. Opting out restores the old fallback.
+### Fixed
+
+- **A supplied title-block `date` is stated on the sheet** (#1585). `TitleBlock`
+  could show only one of `revision`/`date` in its top-right cell and revision
+  won, while draftwright defaults `revision` to `"A"` — so every drawing that
+  set a date lost it, silently, in every version that has had the parameter.
+  Adopting `build123d-drafting-helpers>=0.15.4` gives the date a cell of its
+  own whenever a revision is also set, and `_make_title_block` now names that
+  cell, so the date reaches the PDF text layer and the cell-overflow lint
+  rather than only the rendered geometry.
+
+  `date`, `revision` and `company` are stripped before use, matching what the
+  block itself does: a whitespace-only `company` previously raised `KeyError`
+  from `cell_bbox`, and padded values were measured at a width the sheet never
+  drew. (The DRAWN BY narrowing this entry originally described, and the shared
+  top-right cell it named, are both gone with the ISO 7200 layout above — the
+  date has its own cell in its own row, so it costs DRAWN BY nothing.)
+
+
+### Added
+
+- Report schema v1 now projects one recognition-owned physical requirement ledger shared with
+  completeness. Each bounded requirement identifies its accepted source occurrence(s), final IR
+  owner(s), semantic parameter where applicable, placement/satisfaction state, annotation names,
+  and representation rationale. The roster is evaluated once per report and fails closed when
+  exact same-run evidence or provenance is absent; it does not create a second recognition scan,
+  persistent topology identity, inferred manufacturing intent, or visual change (#1438, ADR 0017
+  Amendment 26).
+
+- Direct CLI rendering now writes `<output>.draftwright.json` by default, after successful visual
+  export, and prints its path with the requested artefacts. `--no-report` is the explicit opt-out;
+  library `Drawing.export()` remains free of report-writing side effects (#1438, ADR 0017
+  Amendment 25).
+
+- Generated Sheet Python now embeds a deterministic, versioned, JSON-compatible snapshot of
+  actionable accepted-occurrence gaps from the exact recognition run that produced its model.
+  The compact snapshot carries only unsupported, deferred, evidence-only, and unexpectedly
+  missing public records; STEP inputs are read once and recognition uses a private copy of the
+  exact bytes named by the embedded basename and SHA-256, with a best-effort final replay-target
+  check before writing. The hash is a later reconciliation boundary, not a lock on the mutable
+  source path. It is explicitly generation-time evidence, never current completeness authority
+  (#1438, ADR 0017 Amendment 24).
+
+- Accepted Plate occurrences now retain exact conditional ownership. Genuine multi-axis plates
+  keep their direct `PlateFeature`; other slabs are absorbed only when released same-run AAG
+  evidence joins them to the exact retained face-level, riser-shoulder, or slot-pattern owner.
+  Multi-owner claims are conjunctive, while whole-envelope, polygonal-boss, disconnected,
+  ambiguous, or malformed cases remain visibly `unexpectedly_missing` rather than borrowing an
+  owner from matching values or traversal order (#1438, ADR 0017 Amendment 23).
+
+- `Drawing.write_report(path)` now atomically writes the versioned report as deterministic,
+  strict UTF-8 JSON without coupling library export to sidecar generation. Report and filesystem
+  failures leave any existing destination untouched, with best-effort temporary cleanup that does
+  not mask the primary error (#1438, ADR 0017 Amendment 22).
+
+- Raw automatic drawings now expose `Drawing.report()`, a strict version-1 JSON-compatible
+  projection of every accepted recognition occurrence, its explicit consumer disposition and
+  report-local final IR owner(s), plus the existing lint summary. The published schema never
+  serializes provider references, topology IDs, object addresses, or transient member positions;
+  it makes requirement coverage an explicit field (subsequently populated by Amendment 26) and
+  calls a clean result `bounded-clear`, not manufacturing-ready. Paths without exact occurrence
+  ownership and raw paths with an unclassified accepted occurrence fail explicitly rather than
+  emit an empty or inferred report. Rendering remains unchanged (#1438, ADR 0017 Amendment 21).
+
+- Accepted `FaceLevel` and `RiserEvidence` occurrences now receive explicit `evidence_only`
+  outcomes. The released records remain shared substrate for the supported correlated height
+  ladder and shoulder projection, not independent inferred features or drawing requirements; no
+  IR, Sheet DSL, renderer, or recogniser API change is introduced (#1438, ADR 0017 Amendment 20).
+
+- Accepted round-boss occurrences now retain the exact existing consumer outcome: a singleton
+  `BossFeature`, one diameter-consolidated representative with transient member lineage, the exact
+  turned-step owner, or the exact groove owner reached through that step or the profile-gate
+  fallback. Ambiguous or missing correspondence fails closed without another scan, persistent ID,
+  generated artefact, or visual change (#1438, ADR 0017 Amendment 19).
+
+- Accepted through-step occurrences now retain either their exact `ThroughStepFeature` or every
+  exact final `StepLevelFeature`, `PlateFeature`, and `EnvelopeFeature` that jointly carries the
+  established compatibility projection. Plate claims are body-local, replacement is atomic, and a
+  missing or ambiguous selected owner fails closed rather than granting partial credit. One
+  occurrence can name multiple run-local IR owners without a second scan, persistent ID, generated
+  artefact, or change to rendering machinery (#1438, ADR 0017 Amendment 18).
+
+- Accepted turned-step occurrences now retain the conversion decision that already gives an
+  ordinary profile band its exact `StepFeature` or absorbs a groove-floor band into the exact
+  same-run `GrooveFeature`. The takeover must be one-to-one in both directions, so an ambiguous
+  groove, an adjacent band inside its positional tolerance, or any missing final owner fails closed
+  as `unexpectedly_missing`, with no second scan, persistent ID, generated artefact, or visual
+  change (#1438, ADR 0017 Amendment 17).
+
+- Accepted channel occurrences now receive an exact conditional consumer outcome. A genuine
+  multi-axis plate channel is `represented` by its `ChannelFeature`; a monolithic Z-depth rebate is
+  `absorbed` only when one retained body-local support carries its floor and both shoulders into the
+  final step ladder. Cross-axis, disconnected, or filtered channels fail closed as
+  `unexpectedly_missing` instead of inheriting an unrelated ladder, without a second scan,
+  persistent ID, generated artefact, or visual change (#1438, ADR 0017 Amendment 16).
+
+- Accepted countersink occurrences now receive an exact run-local outcome. A seat carried by a
+  `HoleRecord` is `absorbed` by that hole's final feature and transient member lineage; an unattached,
+  unowned, or duplicate nested occurrence fails closed without dimension matching, topology IDs, a
+  second scan, duplicate IR requirements, or visual changes (#1438, ADR 0017 Amendment 15).
+
+- Accepted angled-step, passage, prismatic-pocket, oriented-slot, and repeating-radial-profile
+  occurrences now receive exact run-local ownerless outcomes from Draftwright's existing consumer
+  capability declaration: `unsupported`, `deferred`, or report-facing `evidence_only`, with closed
+  reason codes and existing tracking issues. Equal-valued occurrences stay distinct; policy does
+  not invent IR owners, requirements, topology IDs, recognition scans, generated artefacts, or
+  visual output (#1438, ADR 0017 Amendment 14).
+
+- Accepted physical hole, slot, and pocket occurrences now retain explicit run-local ownership
+  when conversion represents a singleton or absorbs members into one same-spec hole group or
+  derived pattern feature. Family-specific reason codes and transient member lineage follow
+  member-specific AP242 splits to the exact final IR objects; missing lineage fails closed rather
+  than retaining a stale owner. Provider-derived patterns do not become invented occurrences, and
+  no report ID, extra recognition scan, generated artefact, or visual output is introduced (#1438,
+  ADR 0017 Amendment 13).
+
+- Raw automatic builds now capture run-local, conversion-time ownership from accepted provider
+  occurrences to their exact Draftwright IR features for unconditional one-to-one adapters. Equal
+  record values, order, coordinates and topology indices cannot create a binding; an omitted direct
+  binding fails closed as unexpectedly missing, while ownership rules not yet implemented remain
+  explicitly unclassified. The experimental read-only `Drawing.recognition_ownership()` exposes
+  the non-serializable ledger; it stays outside the IR/compiler and changes no generated or visual
+  artefact (#1438, ADR 0017 Amendment 12).
+
+- Raw automatic recognition now retains the provider's run-scoped `RecognitionEvidence` beside
+  its exact `RecognitionResult`, exposed through the experimental read-only
+  `Drawing.recognition_evidence()`. Declared drawings acquire the pair lazily on first physical
+  critique; bare and framed results remain evidence-less, and Draftwright never rescans merely to
+  backfill evidence. This is a non-visual reporting foundation and changes no rendered artefact
+  (#1438, ADRs 0015, 0017, 0020).
+
+- Automatic drawings can opt into the provider-owned local recognition frame with
+  `framed_recognition=True`. The returned drawing exposes caller provenance, the exact downstream
+  working solid, frame, and a copied framed/raw/refusal decision. PMI correlation, off-axis pattern
+  placement, compilation, projection and physical lint remain in one coordinate system; typed frame
+  refusal performs one visible top-level raw fallback. Raw remains the rollout default (#1357,
+  ADR 0020).
+
+- Referential and measured Sheet dimensions accept optional `view=` / `side=` placement
+  intent. Supported pairs survive IR/compiler and generated-script round trips, constrain
+  ordinary corridor candidates without exposing page coordinates, and fail clearly when a
+  renderer cannot honour the requested view/strip. Compound hole callouts can therefore be
+  authored to opposite sides of one end view while retaining normal solve and lint behavior
+  (#563).
+- Feature-linked Sheet notes can explicitly satisfy one or more canonical requirement ids with
+  `.note(..., satisfies=(...))`. Placed structured notes retain each requirement's identity in
+  coverage reports and generated scripts without masquerading as dimensions or parsing prose;
+  removing the note restores the ordinary missing/suppressed diagnostics (#1351).
+- `Sheet.from_part(...).take_over(...)` now adopts detected features in place while choosing
+  authored or automatic dimension, principal-view, and derived-view sources explicitly. This
+  preserves detected handles, supports automatic principal views with authored requirements,
+  replaces or suppresses inferred sections without duplicates, and round-trips the adopted view
+  request through generated Sheet code (#1350).
+- Referential Sheet dimensions can preserve an explicit number of decimal places with
+  `sheet.dimension(feature, "parameter.id").format(decimals=n)`. The policy changes only
+  printed display text: numeric reconciliation, tolerances, suppression and dimension
+  provenance continue to use the feature parameter, and generated Sheet scripts round-trip the
+  policy (#1349).
+- New lint code `annotation_ink_overlap`: reports another annotation's line-work —
+  a dimension line, an extension line or a leader shaft — drawn through a label's
+  text, which `annotation_overlap` cannot see because it compares label boxes to
+  each other and to nothing else. Measured as the longest connected stroke path
+  crossing the label's exact polygon (or its axis-aligned box), in millimetres,
+  with a 0.5 mm floor. It scores on legibility, so
+  dense sheets that previously scored clean now score lower — 17 findings on
+  NIST CTC-02 AP203 alone, every sampled one confirmed a real defect by rendering.
+  Repair does not yet act on it (#1333) and the corridor solve does not yet avoid
+  it (#1334) (#1321, #1332).
+
+### Changed
+
+- Updated the exact Quiddity runtime and declared-inspection contract to **0.2.8**, which stops
+  a modelled thread being read as a stack of turned shoulders and coalesces contiguous
+  equal-diameter bands into the one span they physically are. Measured on the repo corpus,
+  NIST CTC-05 is again the only fixture that moves: its ⌀304.8 region was three adjacent
+  rungs and is now one, in both STEP encodings (#1135, quiddity#587).
+
+- Updated the exact Quiddity runtime and declared-inspection contract to **0.2.7**. Unlike the
+  0.2.5 adoption below, this one changes drawings: the provider now requires circumferential
+  cylinder support to establish a turned-profile axis line, so rounded plate corners and parallel
+  offset cylinders no longer form a false stepped shaft. Measured on the repo corpus, NIST CTC-05
+  is the only fixture whose recognised turned profile moves — its two STEP encodings previously
+  disagreed (3 steps and 9 steps for the same solid) and now agree on 5 — and a plate part that
+  previously raised out of `generate_sheet_script` now emits and builds. No family, record schema
+  or adapter changed (#1555, quiddity#586).
+
+- Updated the exact Quiddity runtime and declared-inspection contract to 0.2.5 for the
+  provider's recognition performance improvements.
+
+- Production recognition now consumes published `quiddity==0.2.2`. Unified section-recess
+  occurrences and their same-run patterns feed the existing pocket, channel and blind-slot
+  drawing grammar. Original occurrence ownership, independent completeness and explicit
+  unsupported/refusal outcomes follow that single aggregate (#1471).
+- Machine-readable reports and STEP inspection documents advance to schema v2 because their
+  closed producer object now names `quiddity`. The schema-v1 documents remain available for
+  existing files; readers must select the schema using `schema_version` (#1471).
+- Groove membership now retains Quiddity's body-local profile identity. Generated declarations
+  use `profile_group` to associate coaxial grooves and steps without copying provider keys.
+  Omitted steps do not cause a groove to attach to another body (#1471).
+- Known limitations in Quiddity 0.2.2: some mixed-profile/nested pockets, bore-intersected channels
+  and edge-open recesses, and pockets in cylindrical stock can lose automatic annotations.
+  Refusals remain visible in lint and evidence. Regression cases are retained as strict expected
+  failures while upstream [#536](https://github.com/pzfreo/quiddity/issues/536),
+  [#538](https://github.com/pzfreo/quiddity/issues/538) and
+  [#541](https://github.com/pzfreo/quiddity/issues/541) are open.
+
+- Accepted schema-v3 straight or circular `Blend` chains, including concave occurrences, now cross
+  a dedicated end-to-end consumer path:
+  `BlendFeature`, explicit `Sheet.blend(...)`, executable generated declarations, one
+  compiler-approved `blend.radius` requirement, and a solver-owned grouped `n× R` leader. The
+  discriminated path kind, full canonical direction/normal, circular-path major radius,
+  rolling-ball radius, and material side remain in IR instead of being coerced to legacy Fillet
+  semantics; the provider aggregate remains the sole exact Fillet/Blend precedence authority.
+  Exact public-record/IR validation and completeness follow every accepted occurrence to placed,
+  structured-note, authored-suppressed, dropped, missing, or unverifiable outcomes. Raw and
+  provider-framed arbitrary rigid motions retain straight-path radius meaning; framed recognition
+  also retains circular paths, while raw arbitrarily rotated circular recognition remains an
+  upstream limitation (b123d-recognisers#491) (#1433; ADRs 0011, 0013–0017, 0020).
+
+- The preceding recognition update pinned the immutable `b123d-recognisers` 0.4.14 release and advanced its
+  fail-closed capability and inspection joins. `Blend` schema v3 replaces the old flat cylindrical
+  fields with exact `StraightBlendPath` and `CircularBlendPath` records and admits proved concave
+  and toroidal occurrences; Draftwright preserves those semantics through IR, `Sheet.blend(...)`,
+  generated replay, reporting, drawing, and completeness. Passage nested records advance to schema
+  v2 without changing Draftwright's unsupported passage policy. The provider's previously
+  recognised shallow nonzero `PairedRampStep` keeps its existing
+  two-requirement IR, public `Sheet.paired_ramp_step(...)` declaration, generated-code, compiler
+  callout, solver placement, and completeness meaning (#1430; ADRs 0011, 0013–0017, 0020).
+
+- Rectangular blind slots recognised by `b123d-recognisers` 0.4.10 now cross a dedicated
+  Draftwright semantic path: `RectangularBlindSlotFeature`, the explicit
+  `Sheet.rectangular_blind_slot(...)` word, generated-code and role-specific tolerance round
+  trips, compiler-approved width/capped-run/depth measurements, and one solver-owned
+  `OPEN SLOT … DEEP` leader. Authored subsets retain every approved value with explicit
+  `WIDE`/`LONG`/`DEEP` roles, and leaders target proved cap/side material rather than the open
+  mouth. Native raw and rigidly moved framed builds retain the same three
+  measurement outcomes. A seven-case independently authored corpus now pins all six run/open-side
+  orientations plus an independently sized specimen: each occurrence contributes separate width,
+  capped-run and depth requirements followed by exact structural correspondence and measurement
+  provenance to placed, structured-note, suppressed, dropped, missing or unverifiable outcomes.
+  Removing recognition cannot shrink that 21-requirement benchmark (#1421; ADRs 0011,
+  0013–0017, 0020).
+
+- Round-bottom blind slots recognised by `b123d-recognisers` 0.4.10 now have their own complete
+  Draftwright path rather than borrowing through-slot, pocket or rectangular-blind-slot grammar.
+  `RoundBottomBlindSlotFeature` retains the released run/opening topology and exposes the three
+  independent manufacturing requirements: capped run length, straight bottom-flat width and equal
+  side radius. `Sheet.round_bottom_blind_slot(...)`, generated code, role-specific tolerances and
+  one solver-owned `ROUND-BOTTOM OPEN SLOT …` leader all preserve those exact identities; derived
+  total width and depth are not double-counted. A seven-case independently authored corpus covers
+  all six signed principal orientations plus a separately sized specimen (21 physical requirements),
+  proves raw/framed drawing parity, and follows exact source-to-IR correspondence to placed,
+  structured-note, suppressed, dropped, missing or unverifiable outcomes without labels, views or
+  page coordinates. The released 0.4.10 API is sufficient; no recogniser change is required (#1421;
+  ADRs 0011, 0013–0017, 0020).
+
+- Turned-step completeness is now backed by an independently authored, hash-pinned 11-case STEP
+  corpus with 26 physical outside-diameter bands, 52 parameter checks and 104 downstream checks.
+  Axis line and station preserve body-local occurrence identity across automatic IR, the existing
+  public `Sheet.step(...)` DSL word, executed-and-built generated code, and exact compiler-confirmed
+  length and OD drawing evidence. Grouped equal-length chains retain every adjacent member identity
+  and their exact compiler-approved multiplier and value; native and global OD labels are checked
+  in full. Generated declarations preserve the half-thousandth midpoint of odd-thousandth spans,
+  complete body-local profile ownership is compared at its full published precision, and malformed
+  or surplus records fail closed—including coercible or non-finite groove evidence that controls
+  band ownership. The root turned-step and groove inventories must remain immutable tuples; a
+  mutable or one-shot substitute is snapshotted once only to preserve its observable requirement
+  denominator, then remains unverifiable; a non-iterable substitute retains one explicit aggregate
+  contract outcome rather than masquerading as an empty family. A profile OD may
+  satisfy one uniquely identifiable equivalent largest-band diameter
+  without multiplying across disjoint bodies; and uniquely groove-owned narrow
+  bands are excluded once by a shared fail-closed Groove schema and ownership rule used by lowering,
+  lint and the evidence ledger. That join accounts for the provider's 0.001 mm Groove coordinate
+  quantum against its eight-decimal profile axis without widening either published identity
+  (#1374; ADRs 0010, 0011, 0013, 0015, 0017, 0020).
+
+- Framed recognition now carries all three 0.4.6 step families through their complete existing
+  consumer paths. `CircularBlindStep`, `PairedRampStep` and `ThroughStep` retain the same
+  manufacturing requirements from raw recognition through their public `Sheet` words, generated
+  code, compiler-owned drawing ink and lint; the provider-owned local records and Draftwright IR
+  remain invariant under a combined non-principal rotation and translation. Circular-step
+  validation now derives its bound from each independently published endpoint, centre, radius and
+  depth value's six-significant-figure decimal cell, while outside-cell geometric inconsistencies
+  still fail closed (#1382; ADRs 0011, 0013, 0015, 0017, 0018, 0020).
+
+- Countersink completeness is now backed by an independently authored, hash-pinned seven-case STEP
+  corpus covering functional seats, solid external-cone and deburr negatives, grouped equal seats,
+  and reverse-serialized topology. Detection identity and all four provider measurements are scored
+  separately from the two actual drawing requirements. Exact aggregate object ownership follows
+  each seat once, validates the provider-selected hole semantically, and fails closed on swapped,
+  duplicated, or canonically colliding owners. It then crosses automatic IR, public `Sheet.hole`,
+  generated code, and role-specific confirmed placed diameter/angle ink without recounting the parent bore. The known second seat of a two-sided bore
+  remains explicitly unverifiable rather than hidden or heuristically attached (#1370; ADRs 0010,
+  0013, 0015, 0017).
+
+- Double-D completeness is now backed by an independently authored, hash-pinned 11-case STEP
+  corpus with ten physical profile facts, 50 provider-parameter checks and 40 downstream checks.
+  Full-frame identity preserves disconnected coaxial bodies; exact multiplicity crosses automatic
+  IR, public `Sheet.double_d_bore`, generated code and role-specific compiler-confirmed
+  `⌀major THRU DOUBLE-D across A/F` ink. Round, blind, opposed-blind, malformed, duplicated and
+  foreign-owner mutations fail closed, while the provider's exclusive aggregate ownership keeps
+  each profile from being recounted as an ordinary hole (#1370; ADRs 0010, 0011, 0013, 0015,
+  0017).
+- Updated the immutable `b123d-recognisers` pin and fail-closed inspection/capability joins to
+  0.4.10. All 28 previously consumed family schemas remain unchanged. The two new blind-slot
+  families were adopted fail-closed under #1421 and remain visible as unscored recognised
+  inventories until each gains independent completeness evidence; Draftwright does not coerce
+  their depth, terminal/open-end, or round-floor evidence into its ordinary through-slot grammar.
+  The provider's additive evidence API is not consumed by
+  this bounded adoption slice. Recognition output does change under the unchanged schemas: the
+  release closes slot-depth, subdivided paired-ramp/AngledStep, and noisy stubby-pocket gaps; fixes
+  Double-D/Hole ownership, external-cone countersink false positives, Plate tie covariance, and
+  turned-step translation covariance; and can move an edge-open rectangular recess from `Pocket`
+  to the dedicated blind-slot inventory. Draftwright regression tests pin the consumer-visible
+  ownership changes while the provider release suite owns its private predicates
+  ([provider release](https://github.com/pzfreo/b123d-recognisers/releases/tag/v0.4.10); ADRs 0013,
+  0015, 0017, 0020).
+
+- Principal-profile lint now correlates its independent extremal-face evidence with the cached
+  public `DoubleDBore` aggregate instead of importing provider-private profile helpers. Exact
+  axis/span, end-plane, centre, radius, A/F, chord, arc, and flat-direction checks fail closed;
+  occurrence mouths are consumed with matching multiplicity, and malformed records cannot earn
+  coverage. The exported low-level check now requires the aggregate projection explicitly so an
+  omitted inventory fails loudly; lint still owns its physical denominator and never reruns
+  recognition (#1411, ADRs 0013, 0015, 0017).
+
+- Whole polygonal-stock completeness is now backed by an independently authored, hash-pinned
+  13-case STEP corpus covering all principal axes, in-plane rotation, translation, topology
+  order and circular/rectangular/irregular/octagonal/attached/recessed/compound negatives. The
+  provider's exact six-side stock boundary, A/F, axial length and coupled physical flat supports
+  are scored separately from four downstream boundaries. A fail-closed two-requirement ledger
+  follows A/F and stock length through automatic IR, public `Sheet.polygonal_stock`, generated
+  code and solver-placed semantic ink; malformed or duplicate correspondence cannot borrow
+  credit from another support ring (#1371).
+
+- Multi-plate slab completeness is now backed by an independently authored, hash-pinned 11-case
+  STEP corpus covering all principal orientations, additive/cut channels, equal compound
+  occurrences, topology order and envelope/detached/thick/rotational negatives. A fail-closed
+  one-requirement ledger follows each body-local `thickness.length` through automatic IR, public
+  `Sheet.plate`, generated code and solver-placed Dimension provenance; the compiler-derived
+  opposite U-channel wall is represented explicitly without inventing duplicate ink. The corpus
+  owns 20 physical occurrences and cannot shrink when provider output is weakened or removed.
+  Exact step-level, slot-pattern, polygonal-boss and envelope ownership prevents derived material
+  spans from becoming duplicate Plate requirements while failing closed when that ownership cannot
+  be established (#1373).
+
+- Polygonal-boss completeness is now backed by an independently authored, hash-pinned STEP
+  corpus covering all three principal axes, in-plane rotation, equal compound occurrences,
+  topology order, and recess/detached/stock/other-protrusion negatives. Detection, the provider's
+  six-side schema invariant, across-flats, height, physical flat supports, and four downstream
+  boundaries are observed; a
+  two-requirement fail-closed ledger follows across-flats and boss height through the existing
+  IR/compiler/shared-solve path. The corpus owns the independent evaluation denominator; runtime
+  drawing lint remains explicitly denominated by the recognition inventory (#1372).
+
+- Updated the immutable `b123d-recognisers` pin and fail-closed manifest joins to 0.4.9. The
+  provider's body-local FaceLevel, RiserEvidence, and TurnedProfile schemas now cross an owned
+  prepared-frame boundary that normalizes, reuses one cylinder substrate, classifies locally,
+  and preserves the exact working solid/result pair. Every frame gauge and typed refusal has an
+  explicit conservative policy; refusal never hides a raw retry. Multiple physical turned
+  profiles fail explicitly where a singular adapter is required, while the unchanged raw
+  production route preserves its historical no-global-profile compound behavior with a visible
+  #1357 deferral. Production drawing remains on the explicit raw boundary until rigid-motion,
+  PMI, platform, and canary activation evidence is complete (ADR 0020, #1357).
+
+- Updated the immutable `b123d-recognisers` pin and lock to 0.4.8. Draftwright now names
+  its existing caller/world-coordinate aggregate boundary explicitly with
+  `build_raw_recognition_result`; framed recognition remains disabled until #1357's
+  frame-to-IR adapter is reviewed end to end. Installed-wheel tests exercise the public
+  framed fixes for rectangular pads, polygonal bosses, and nested multi-body plates without
+  making that API the production path (#1392).
+
+- Rectangular pads now consume `RaisedPad` schema v2 without assuming +Z. The IR, public
+  `Sheet.pad`, generated code, compiler, semantic-view planner, shared placement solve, and
+  lint correspondence preserve the exact footprint, terminal-to-attachment height, two
+  in-plane locations, axis, direction, and world bounds for all six signed principal
+  orientations. Every height uses one solver-placed `… HIGH` leader with stable measurement
+  identity; side-normal pad footprints no longer leak into the unrelated Z-level grammar
+  (#1392).
+
+- Rectangular-pad completeness is now independently evidenced by a hash-pinned 12-case STEP
+  corpus covering all six signed principal orientations, equal compound occurrences, topology
+  order, and ledge/staircase/detached-body negatives. Detection, width/length/height fidelity,
+  and four downstream boundaries are scored separately; a five-requirement fail-closed ledger
+  follows both footprint sizes, height and directional location evidence through the shared
+  compiler and placement solve (#1372).
+
+- Circular blind steps recognised by `b123d-recognisers` now have a complete consumer path.
+  The oriented terminal-to-open centreline and transverse quarter-arc lower without rescanning
+  to an explicit-only IR/Sheet feature; generated code preserves the full correspondence.
+  Radius and stopped depth remain independently addressable and tolerance-aware while sharing
+  one solver-placed end-view `R… × … DEEP` leader, and a fail-closed two-requirement ledger
+  follows each outcome without parsing labels or guessing by order (#1382).
+
+- Paired-ramp steps recognised by `b123d-recognisers` now have one complete consumer path:
+  exact IR conversion, explicit `Sheet.paired_ramp_step(...)` declaration, generated-code
+  round trip, end-view solver-placed `2× angle × run RUN` leader, independent angle/run
+  tolerances and authored suppression, and a two-requirement fail-closed completeness ledger
+  (#1382).
+
+- Rectangular through steps recognised by `b123d-recognisers` now lower without rescanning to
+  an explicit-only IR/Sheet feature. Automatic and explicit lowering support every principal
+  axis; an X/Y occurrence stays with the established face-level and shoulder/plate grammar only
+  when that grammar proves both exact leg intervals, avoiding both silent loss and duplicate
+  dimensions. The two
+  orthogonal open-section legs are independently addressable, tolerance-aware, solver-placed in
+  the end view, preserved by generated code and scored by a fail-closed two-requirement
+  completeness ledger. The through extent stays owned by the envelope (#1382).
+
+- Hole completeness is now supported by a versioned, independently authored STEP corpus rather
+  than a Draftwright-derived census. Detection, parameter fidelity and downstream usefulness stay
+  separate, while the IR adapter, public Sheet declaration, executed generated code and placed
+  drawing are each observed through their real seam. Mutations prove that a missing provider hole
+  or a broken downstream boundary lowers the relevant evidence even when internal state remains
+  self-consistent (#1369).
+- Hole-pattern completeness is now supported by a separate independently authored STEP
+  arrangement corpus covering grids, bolt circles, linear arrays, false-positive controls,
+  compound cases and topology-order variants. It scores one grouping fact per pattern while the
+  hole corpus remains the sole owner of member diameter/depth/location requirements, and follows
+  each group through the IR, public Sheet declaration, generated code and placed pattern
+  dimensions (#1370).
+- Flat completeness is now supported by an independently authored STEP corpus covering lone,
+  opposed, slanted, parallel and coaxial stock plus negative and topology-order controls. It
+  counts one physical across-flats requirement per stock line and span, then observes automatic
+  IR, public `Sheet.flat`, executed generated code and placed measurement provenance independently
+  (#1371).
+- Lone-pocket completeness is now supported by an independently authored STEP corpus covering
+  off-centre, edge-anchored, equal, opposed-opening, side-opening, compound, ownership-negative
+  and topology-order cases. Provider `open_sign` now survives the IR, declaration and generated
+  code boundaries; pocket sizes and both physical location ordinates are followed through exact
+  semantic placement evidence without counting pocket-pattern members twice (#1372).
+- Pocket-pattern completeness is now supported by a separate independently authored STEP corpus
+  covering axis-aligned and 30° linear/grid arrangements, threshold and unequal-spacing
+  negatives, a compound underside pattern and reverse-serialized topology variants. One grouped
+  count/size/lattice/location requirement reaches automatic IR, public `Sheet.pocket_pattern`,
+  generated code and exact placed ink—including approved pitch tolerances—without recounting its
+  member pockets. Diagonal pocket-pattern dimensions use exact label-ink clearance instead of a
+  false-blocking hull; generated code preserves the member opening side and grouped callouts
+  retain structured count provenance (#1372).
+- Groove completeness is now supported by an independently authored STEP corpus covering a
+  monotonic-step negative, all principal axes, a narrow circlip ownership contest, equal grooves
+  on separate bodies and reversed Boolean order. Each physical groove's width and floor diameter
+  pass independently through automatic IR, public `Sheet.groove`, executed generated code and
+  one exact compiler-backed `WIDE × ø` callout; mutations prove that recognition, identity,
+  parameters, provenance and printed syntax cannot disappear without lowering evidence (#1372).
+- Chamfer completeness is now supported by an independently authored STEP corpus covering all
+  principal axes, equal and asymmetric planar bevels, conical turned treatments, equal features
+  on separate bodies, topology-order variants, a plain negative and the AngledStep ownership
+  boundary. Each physical bevel is followed through exact IR, public `Sheet.chamfer`, executed
+  generated code, compiler measurement provenance, callout syntax and its physical leader target;
+  equal requirements may share ink without collapsing occurrence identity (#1374).
+- Fillet completeness is now supported by an independently authored STEP corpus covering all
+  principal axes, repeated equal rounds, turned toroidal treatments, separate bodies,
+  topology-order variants, a plain negative and the CircularBlindStep ownership boundary. Each
+  physical round is followed through exact IR, public `Sheet.fillet`, executed generated code,
+  compiler measurement provenance, `R`/`n× R` syntax and its physical leader target; grouped ink
+  retains every rounded edge's measurement identity (#1374).
+
+- Recognised angled blind steps now fail visibly instead of remaining an undecided inventory.
+  Each aggregate-reconciled `AngledStep` produces an actionable warning and one `unsupported`
+  completeness outcome. A frozen STEP corpus fixture pins the AngledStep/Chamfer ownership
+  boundary, while Draftwright deliberately invents no angle/run, IR, Sheet, generated-code,
+  section/detail, or annotation semantics (#1247).
+
+- Aggregate-reconciled `PrismaticPocket` occurrences now fail visibly instead of
+  remaining an unscored inventory. Each produces an actionable warning and one `unsupported`
+  completeness outcome, without repeating the provider's cross-family reconciliation or
+  inventing width/length, across-flats, IR, Sheet, generated-code, or annotation semantics (#1246).
+
+- Recognised prismatic passages now fail visibly instead of disappearing into an unscored
+  inventory. Each authoritative `SectionPassage` produces an actionable warning and an
+  `unsupported` completeness outcome; the legacy `Passage` compatibility projection is never
+  double-counted. Draftwright retains the family-level unsupported disposition rather than
+  claiming a regular-polygon callout represents the provider's complete line/arc schema (#1245).
+
+- Declared Double-D bore, countersink, chamfer, fillet, and groove geometry reads now use the
+  stable `b123d_recognisers.inspection` namespace from the exact 0.4.6 wheel. A separate
+  Draftwright-owned format-1 contract fails closed on changes to the consumed signatures, result
+  schemas, units, bevel rejection reasons, or cylindrical surface layout (#1362).
+
+- Updated the exact `b123d-recognisers` production lock to 0.4.6. At adoption time, circular
+  blind, paired ramp, and through-step inventories were carried from the one aggregate and
+  surfaced as unscored completeness evidence pending the #1382 semantic slices now recorded
+  above. Rectangular-grid pitch placement also uses a
+  deterministic outer witness line across plan, front, and side views despite the release's
+  coordinate-rounding differences.
+
+- `annotation_overlap` now reports the whole defect when the same pair also draws
+  line-work through one of the two labels: the message names the crosser, the label
+  crossed and the millimetres, and replaces the "use `label_offset_x`" remedy, which
+  is the advice that case contradicts. `LintIssue.suggestion` is `None` for those
+  issues rather than a snippet that re-places the text (#1321, #1332).
+
+- Recogniser integration now uses only the exact `b123d-recognisers` dependency in
+  `pyproject.toml` and `uv.lock`. The separate release-evidence record, dependency-bump workflow,
+  candidate-checkout harness, and release-time capability-version rewrites have been removed;
+  ordinary compatibility tests validate the installed pinned package.
+- Automatic composition now measures the next larger preferred scale when a conservative
+  crowded-detail reservation leaves a sparse sheet. A candidate wins only on the same paper
+  size, without the recovery detail, structural errors, or required-placement loss. GRM-04 now
+  uses A4 at 5:1 instead of 2:1, re-homes its shared side-hole Y location into the plan-view
+  corridor, and retains both the Ø2.4 diameter and location requirements (#1155).
+- Closed the remaining dense-sheet leader cases as deliberate Policy-B behavior: exhaustive
+  route sweeps prove there is no route clear of both material and committed sheet ink, so the
+  required callout remains with an explicit crossing diagnostic instead of being dropped. A
+  fast synthetic release gate complements the retained exhaustive corpus characterization
+  (#1187).
+
+### Fixed
+
+- Overall height and envelope width remain dimensioned when approved turned-step lengths
+  cover only part of the respective extent. Redundancy suppression now requires adjoining
+  measurements on one profile to reach both overall endpoints; STC618 retains its missing
+  62 mm height (#1509).
+- Non-1:1 orthographic and isometric projection now scales extracted solids about the world
+  origin, matching the view-coordinate mapper and solver zones. Solids carrying a non-zero
+  build123d `Location` no longer shift their silhouette away from dimensions and callouts.
+- Notes and generic table cells now render U+2300 `⌀` with the established supported drafting
+  glyph instead of IBM Plex Mono's missing-glyph square. The substitution happens at the one
+  shared measurement/render seam, so source rows remain unchanged and table sizing stays exact
+  (#1275).
+- Automatic turned-part composition now gives true shoulder coverage priority over optional
+  pictorial content. GRM-03 replans to 5:1 and renders its 0.5, 2, 3, and 18 mm axial lengths;
+  `axial_length_missing` disappears because the dimensions are present, not because the
+  diagnostic was suppressed (#443).
+
+## v0.4.10 — 2026-08-22
+
+**A view-set-aware planning and turned-edge-treatment patch.** Authored view constraints now
+compile into an explicit resolved plan, automatic drawings fail closed when required content is
+lost, AP242 hole tolerances reach their semantic feature owners, and the exact
+`b123d-recognisers` 0.3.0 release supplies turned chamfer and fillet evidence through truthful
+profile-view callouts.
+
+### Added
+
+- `Sheet` can now author a complete or augmenting semantic view set with `authored_views()`,
+  `auto_views()`, `view()`, `section_view()`, `detail_view()`, and their `add_*` forms. Fluent
+  view handles declare relationships, alignment, scale, and whole-view pins without exposing
+  annotation coordinates; unsupported or infeasible constraints fail at their declaration
+  source instead of being silently relaxed. `sheet.view_constraints` exposes the immutable
+  request and `drawing.view_plan` the distinct resolved result (#1260).
+- Drawings now retain their resolved view topology, per-view requirement coverage, and
+  arrangement decision as inspectable values. The engine can build and semantically compare
+  alternative arrangements and requested view sets while keeping projection and page placement
+  derived from the same plan (#1130).
+
+### Changed
+
+- Dimension planning is now aware of the selected principal view set. Requirements supported by
+  another planned view are re-homed before projection; an unsupported set raises one aggregated
+  `ViewPlanIncomplete` naming every unmet requirement rather than failing late or producing a
+  plausible-looking drawing with missing dimensions (#1259).
+- Three sparse corpus drawings now use a `stacked-iso` arrangement and fit A4 instead of A3 at
+  the same scale, with identical annotations, labels, and build issues. An alternative may only
+  compact the scale already selected by the default arrangement and must survive a real compile
+  without semantic loss; otherwise the default is retained (#1130).
+- Updated the exact `b123d-recognisers` production lock from 0.2.9 to 0.3.0.
+  The immutable PyPI wheel/sdist hashes and capability-manifest digest `6645def94cbabaca7b9676fdb029b02e467266d1f020ea536e903b7d86b52685` are
+  recorded in `.github/recogniser-release.json`; focused compatibility evidence and the
+  normal consumer gates run on the generated PR.
+- Accepted the additive Chamfer and Fillet schema-2 recogniser contract alongside schema 1 for
+  the `b123d-recognisers` 0.3.0 cutover; all other record schemas remain fail-closed at version 1
+  (#1276).
+
+- Updated the exact `b123d-recognisers` production lock from 0.2.6 to 0.2.9.
+  The immutable PyPI wheel/sdist hashes and capability-manifest digest `805876905d548fd0ef1301e2cf24427a86d34bc2e592b8d2d0cb43db94e1d3bd` are
+  recorded in `.github/recogniser-release.json`; focused compatibility evidence and the
+  normal consumer gates run on the generated PR. The aggregate now inventories chamfers and
+  fillets on turned parts, and the fixed coaxial-post slot regression retires its downstream
+  strict xfail.
+
+### Fixed
+
+- Automatic builds now apply the same required-outcome gate as explicit-scale builds. A settled
+  drawing that lost required placement outcomes records `plan_incomplete`, reports an
+  `incomplete` scale decision, preserves measurement/hole/source provenance, and fails lint
+  rather than presenting a diagnostic drawing as complete (#1250).
+
+- Supported AP242 diameter tolerances now correlate with their canonical hole or pattern owner at
+  the completed IR boundary and round-trip through emitted `Sheet` declarations with source
+  provenance. Unmatched, ambiguous, conflicting, partial-pattern, and unsupported requirements
+  fail closed to provenance-preserving measured dimensions rather than being duplicated or lost
+  (#1116).
+
+- A declared step shoulder coincident with its datum is now pruned before rendering instead of
+  sending identical endpoints to the border-dimension helper and raising `ValueError`. Valid
+  siblings remain, while the omission is recorded in `Drawing.suppressions()` and reported as
+  `step_position_coincident_with_datum` (#924).
+
+- Turned chamfer and fillet leaders now read in a shaft profile view and land on their physical
+  edge-treatment sites instead of collapsing axial stations in the end view. Cylindrical
+  feature notes, knurl callouts, and surface finishes likewise derive a model-space surface
+  target rather than pointing at the shaft axis; all continue through the shared placement
+  solves, with X/Y/Z parity and declared Sheet round-trip (#1276).
+
+- Toroidal fillets returned for turned parts by the shared recognition inventory now lower into
+  the existing one-feature-per-round IR and render through the common leader solve. Equal radii
+  share one `n× R` callout carrying every physical fillet's measurement provenance, and emitted
+  Sheet programs preserve the full inventory without rescanning geometry (#1281).
+
+- Turned chamfers returned by the shared recognition inventory now lower into the same
+  one-feature-per-edge-treatment IR as prismatic chamfers. Equal specifications render as one
+  `n×` callout carrying every member's measurement provenance; different chamfer forms do not
+  collapse merely because their first legs match, and distinct tolerances split into truthful
+  requirements instead of being lost or applied to untoleranced siblings (#1254).
+
+- Recogniser dependency automation now starts a fresh `Unreleased` section when the previous
+  Draftwright release has finalized the changelog, so a published recogniser can enter through
+  the documented immutable update workflow instead of failing before it opens a PR (#1279).
+
+## v0.4.9 — 2026-08-21
+
+### Added
+
+- Lint now verifies that an annotation claiming to carry a measurement actually renders it
+  (#1217). Coverage checks read provenance riders the annotations carry about themselves and
+  believed them; nothing checked the claim against what the annotation draws. Four new codes —
+  `claimed_value_absent`, `claimed_measurement_not_compiled`, and two that report an
+  unverifiable claim rather than passing it — and `Drawing.add_table` now keeps the rows it
+  draws so a table's claims are readable back. Measured over all 21 STEP fixtures: 1,081
+  claims, 1,078 confirmed, and three findings — all instances of one genuine provenance
+  defect, see #1219.
+
+- `lint_summary()["quality"]` gains a fourth component, `fidelity`: whether what the drawing
+  says is TRUE. The existing three ask whether required content landed (completeness),
+  whether there is too much of it (restraint), and whether a reader can make it out
+  (legibility) — none asked whether it is correct, so a drawing labelled 99 over a 16 mm
+  path reported a perfect score on every component available to it. Also covers a callout
+  drawn for a feature the part does not have, and a gear data table stating a tooth count
+  the geometry contradicts — the last of which reports `passed: True` and zero errors, so
+  the only existing channel that names it is `by_code`. Completeness, restraint and fidelity
+  report `available: False` with a `reason` rather than a flattering number when they have no
+  evidence, and every lint code the engine emits is classified onto exactly one component,
+  an explicit unscored register, or a register of codes whose component depends on the
+  issue's `outcome_stage`. `quality["unscored"]` reports the findings that reached no
+  component at all and flags any whose classification nobody made — `section_dropped` was
+  found scoring on nothing while a comment beside its emission said the opposite, and eleven
+  more `*_dropped` codes reach the engine as `drop_code` data with the same shape (#1176).
+
+### Changed
+
+- Updated the exact `b123d-recognisers` production lock from 0.2.4 to 0.2.6.
+  The immutable PyPI wheel/sdist hashes and capability-manifest digest `aa52bc04216cd1eccc05796d2262547c7f476b80a3ec77334a77b9d8e3241a58` are
+  recorded in `.github/recogniser-release.json`; focused compatibility evidence and the
+  normal consumer gates run on the generated PR.
+- `HoleRequirementOutcome` gains `members` and `features`, `ClaimOutcome` gains `measurement`,
+  and `linting.hole_coverage.canonical_hole_sites` is public (#1217). A consumer attributing a
+  requirement outcome to a specific recognised hole needs the evidence the outcome accounts
+  for, and needs to key into the canonical space the ledger publishes it in — a through hole's
+  own-axis coordinate is zeroed there.
+- The STEP-analysis benchmark's `drawing_consumer` boundary now consumes
+  `hole_requirement_outcomes` instead of a second correspondence implementation of its own
+  (#1206), and follows the ledger's pointer through `linting.evidence` rather than trusting it.
+  Holes the ledger reports `unverifiable` now score `unknown` where the duplicate credited them
+  by guessing: 26 holes, five on each CTC-03 fixture and eight on each CTC-04 fixture. The
+  five-case evaluation corpus is unchanged, and one lint message on CTC-03 now reports its
+  coordinates in the ledger's own canonical space.
+
+- `Drawing.suppressions()` rows gain a `conveyed_by` key: the dimension that states a
+  withheld measurement instead, or `None` when nothing takes the fact over (#1154). It is not
+  a synonym for the existing `authored` flag — that says whose decision it was, this says
+  where the measurement went.
+
+- Updated the exact `b123d-recognisers` production lock from 0.2.2 to 0.2.4.
+  The immutable PyPI wheel/sdist hashes and capability-manifest digest `d0023d30e583c03abc55f09dfeaaa56fddf56334afa0417818480b2fa2ce3f0f` are
+  recorded in `.github/recogniser-release.json`; focused compatibility evidence and the
+  normal consumer gates run on the generated PR.
+- Updated the exact `b123d-recognisers` production lock from 0.2.1 to 0.2.2.
+  The immutable PyPI wheel/sdist hashes and capability-manifest digest `7985befc8572bfe6d1c0805dfdf690d74a5a4a5f14d5988cc29c366315ff04ca` are
+  recorded in `.github/recogniser-release.json`; focused compatibility evidence and the
+  normal consumer gates run on the generated PR.
+
+### Fixed
+
+- An authored tolerance the compiler approved now reaches the sheet wherever the measurement
+  is drawn (#1215, #1216). `build123d_drafting` resolves a dimension's text as
+  `label if label is not None else _number_with_units(measured, tolerance)`, and every
+  dimension this engine emits passes a label — so a forwarded `tolerance=` rendered nothing
+  while type-checking and reading back correctly from `Dimension.label`. Ten render paths
+  discarded the band that way: envelope extents, the height ladder, the turned-step chain, two
+  public `Drawing` verbs, the deferred corridor route, the detail-view redraw, the short-rise
+  escape, step shoulders, pattern pitch, compound-callout recess terms, grid pitch, and both
+  hole tables. A general guard now decorates every parameter of every feature across a corpus,
+  through both `decorations=` key shapes, and joins what the compiler approved against what the
+  claiming annotation renders.
+
+- Collapsed `N×` marks no longer state a tolerance their members do not share (#1216). A
+  pattern's pitch claimed the authored band of every gap although the recogniser admits 2%
+  jitter — holes at −30, −10, 10.3, 30 printed `3× 20 ±0.1` over a 20.3 mm gap — and fillet and
+  flat collapses took the "first-authored" band, so one of four fillets at ±0.1 printed
+  `4× R5 ±0.1`. Equality is now judged at the drawn precision, and a withheld band is reported
+  (`pattern_pitch_tolerance_withheld`, `collapsed_tolerance_withheld`) rather than dropped
+  silently.
+
+- An approved dimension that cannot be placed is now reported instead of vanishing (#1216).
+  The overall extents' corridor candidate carried a no-op drop handler — the only drop in the
+  engine that recorded nothing — and the step ladder's legibility floor skipped rungs without
+  incrementing any counter. Both now record against the measurement they concern, and a
+  withholding is withdrawn if a later pass draws the measurement after all.
+
+- A feature leader filling a view's below corridor no longer costs the part its overall extent
+  (#1236). The mandatory width or depth now retries on the view's above strip after every
+  corridor has drained, the fallthrough starved slot and plate dimensions already use. Live on
+  three NIST CTC fixtures, which drew no overall width.
+
+- The iso view and the above strips are no longer invisible to each other (#1240). Strip
+  placement never saw the iso, and the iso fit re-scaled after annotation without consulting
+  what had been placed — so growth could invade dimension ink, which it did on CTC-01. The fit
+  is now capped by measured re-projection rather than by a linear model of it: the projected
+  bbox translates as it scales, by up to 7.35 mm on a part carrying a Location, which no
+  prediction about the page centre can express.
+
+
+- A part whose feature-local extent runs between the same two faces as an overall extent
+  is no longer dimensioned twice (#1154). GRM-04's drive plate printed `4.5` twice — once
+  as the hub's height and once as the overall thickness — because two detected records
+  claim one physical fact. The overall extent keeps it; the feature-local one is withheld
+  and records which dimension the reader finds it on. Reconciliation is by exact
+  support-plane coincidence, never by value: a square part's two equal extents run between
+  different faces and remain two facts (#997). A measurement is handed over only when the
+  extent receiving it is actually drawn — by the planner's rules, by the compiler's
+  overall-height rules, and by an authored set — and never when the yielding dimension
+  carries a tolerance, since a toleranced dimension and an untoleranced one are not the same
+  requirement. (A tolerance on the *receiving* extent is the same requirement on the same two
+  faces, and does not prevent the consolidation.)
+
+## v0.4.8 — 2026-08-16
+
+### Fixed
+
+- Compatible automatic/deferred sparse ordinary side/plan hole and post-drain
+  machined-feature callouts now share one bounded same-view leader inventory
+  (#1166). Pattern, profiled-bore, and dense table-eligible hole queues retain
+  their specialised established paths because their winners affect downstream
+  furniture/table semantics or await #798 silhouette-aware routing. The solve
+  preserves required/priority-ranked evidence, rejects new-leader crossings,
+  checks complete leader ink against actual-width fixed dimension/witness,
+  arrow, label, and centre-furniture components (including rendered arrows on
+  shifted dimensions and component-local dashed circular centre furniture);
+  fixed residual faces and the selected survivor both require each tessellation
+  triangle to fit one same analytical component rather than merely the union,
+  plus exact continuous OCC containment for curved faces; residual lowering
+  also retains filled datum/GD&T faces absent from segment metadata; if candidate
+  or selected-survivor construction raises, or a survivor fails validation, the
+  bad alternative fails closed and the canonical lazy producer tail is replayed
+  instead of falsely dropping an otherwise valid callout; an all-invalid tail is
+  reported as validation-stage geometry failure rather than no clear room (so
+  scale fallback does not retry a non-spatial helper defect), while compiler
+  invariant violations such as a missing semantic callout label remain loud;
+  only explicitly marked global turning axes permit an arrow-local leader-tip
+  attachment; near-collinear shaft travel and ownerless section/cutting lines
+  remain fixed ink. The solve prefers
+  clear routes over explicit Policy-B fallbacks, and then minimises real leader
+  length. Provisional sections cannot veto that primary result: one separately
+  bounded refinement may choose an equally complete leader inventory that clears
+  their exact components, after which the final section performs one bounded
+  end-symbol repair and yields if exact ink still conflicts. The selected OCC survivor is validated against the same analytical
+  ink contract. A retained fixed-ink crossing now emits the machine-readable
+  `feature_leader_crossing` info finding instead of appearing lint-clean. Candidate,
+  fixed-component lowering, fixed-obstacle-probe, pair, and exact-search work is
+  bounded; the fixed inventory is lowered once and reused with the producer's
+  legacy lazy result as the floor, including state-cap exhaustion; page,
+  own-silhouette, and title-reservation constraints remain hard in every fallback.
+  A replay that exhausts exact fixed-probe classification is retained visibly as
+  `feature_leader_fixed_ink_unverified` rather than rescanning past the cap, and
+  that uncertainty participates in the legibility quality inventory.
+  Trace output preserves an abandoned fully admitted inventory when state search
+  exhausts, records the producer-floor replay separately, and names exact committed/provisional components,
+  refinement status, and competing blockers; live
+  single-callout edits remain immediate.
+
+## v0.4.7 — 2026-08-16
+
+### Added
+
+- Recogniser co-development now has a one-command, wheel-based two-checkout check; an exact
+  registry-release evidence record; automated dependency PR preparation; and documented landing,
+  compatibility, deprecation, rollback, ownership, and CI-budget rules. Dependency PRs retain the
+  canonical coverage/Codecov gate while avoiding four duplicate platform suites and the slow tier
+  (#1170).
+- A versioned, independently authored STEP-analysis corpus and evaluator now report detection
+  recall/false positives, parameter fidelity, and downstream usefulness as separate evidence
+  layers. Hash-pinned positive, negative, ambiguous, compound, and topology-order-variant fixtures,
+  per-layer diagnostics, and an anti-self-validation mutation prevent recogniser output from
+  defining its own denominator (#1169).
+- Draftwright now validates the released `b123d-recognisers` 0.2.0 capability manifest against
+  an exhaustive consumer-owned declaration of IR, DSL, generated-code, drawing, completeness,
+  and documentation states. Unknown families, stale implementations, record-schema drift, and
+  unevidenced state transitions fail closed; intentionally geometry-only repeating-profile
+  evidence remains usable without acquiring invented drafting semantics.
+
+### Changed
+
+- Updated the exact `b123d-recognisers` production lock from 0.2.0 to 0.2.1.
+  The immutable PyPI wheel/sdist hashes and capability-manifest digest
+  `383a196329ed6a3d1bf88f18ae4ff75613c032f35d40f2b53302b6744da4d19e` are
+  recorded in `.github/recogniser-release.json`; focused compatibility evidence and the
+  normal consumer gates run on the generated PR.
+
+### Fixed
+
+- Stable release and TestPyPI snapshot builds now update package metadata and the independently
+  validated Draftwright consumer-contract identity atomically. This prevents stripped `.dev0`
+  and numbered development artifacts from failing their own capability-contract validation.
+- Installed wheels validate every portable recogniser-contract property without requiring the
+  repository-only behavior-test files whose existence remains a strict source/CI gate. This keeps
+  the packaged STEP-analysis evaluator usable without weakening evidence checks in development.
+- Post-release development-version bumps now update the consumer-contract identity on a branch,
+  open a protected PR, and dispatch the narrow maintenance gate instead of attempting a rejected
+  direct push to protected `main` (#1170).
+- Machined-feature leader callouts now collect all clear alternatives within each
+  post-drain pass and use a bounded maximum-cardinality, minimum-leader-length
+  assignment instead of taking corners greedily in feature order (#740). If a
+  deterministic alternative-count, pass-size, candidate-pair, or search budget is
+  reached, placement retains the legacy greedy result or a result no worse than it
+  rather than losing additional callouts. Pre-drain diameter clearance ranking and
+  its Policy-B fallback remain unchanged.
+
+## v0.4.6 — 2026-08-15
+
+### Changed
+
+- **Geometry recognition is supplied by the standalone Apache-2.0
+  [`b123d-recognisers`](https://github.com/pzfreo/b123d-recognisers) package.** The cutover
+  pins the published stable release `v0.1.0` (built from commit
+  `9e622716c14d491729b5191aa6e5d8351c982a51`), removes the
+  duplicate embedded implementation, and keeps only identity-preserving
+  `draftwright.recognition` / `draftwright.score` compatibility re-exports until 0.6.0.
+  Draftwright continues to own the one-result build/lint cache, record→IR conversion, drafting
+  policy, placement, and critique. The package migration is manifest-pinned and golden-backed;
+  its sole normalization makes an exact dominant-axis tie deterministic across operating
+  systems without changing feature policy.
+
+### Fixed
+
+- Hole-callout obstacle checks now account for the rendered leader shaft width and
+  tip-local arrowhead flare instead of treating the leader as a zero-width centreline
+  (#367). The analytical check remains geometry-only and does not over-reserve the
+  arrowhead's clearance along the whole shaft.
+- Exact X/Z slanted flats retain their established lint-clean side-view callout after the
+  recogniser package's cross-platform semantic-axis normalization. The normalized recognition
+  record remains unchanged; `FlatFeature.presentation_axis` carries the downstream drawing-view
+  choice, preventing a leader/silhouette crossing without duplicating recognition.
+
+## v0.4.5 — 2026-08-15
+
+**A focused technical-drawing correctness patch.** Cross-drilled hole locations now remain
+complete and ordered, datum-starting pocket annotations avoid redundant dimensions, and repeated
+equal-radius fillets collapse into one manufacturing callout.
+
+### Fixed
+
+- Side-drilled hole heights that relocate around callout leaders now join the alternate
+  view's shared dimension ladder before placement, preserving every height and keeping
+  feature locations inside the overall dimension. Datum-starting open pockets no longer
+  print a redundant half-length centre offset, and equal-radius fillets across different
+  edge axes collapse into one counted ``n× R`` callout.
+
+## v0.4.4 — 2026-08-15
+
+**A semantic-completeness and fail-closed layout patch release.** Hole requirements and drawing
+quality are now machine-readable, hole-table replacement is transactional, explicit scales
+preserve required output, notes tables use all viable sheet space, and gear declarations,
+recognition evidence, and callout metadata improve diagnostic trust.
+
+### Added
+
+- **External spur gears can be declared through the `Sheet` façade** (#1086). The declaration
+  records the correlated ISO 1328-1 manufacturing requirements, renders a structured gear-data
+  table, and survives script emission, declaration, export, and lint without inferring gear data
+  that the caller did not supply.
+
+- **Recognition carries serialisable repeating radial-profile evidence** (#1087). Exact
+  rotational whole-wire proof can reconcile a declared gear with its physical profile while
+  preserving ambiguity across separate bodies and without inventing module, pressure angle, or
+  other manufacturing intent.
+
+- **Explicit drawing scales now protect required annotation completeness** (#1146).
+  `build_drawing(...)`, `make_drawing(...)`, `Sheet(...)`, and the CLI accept
+  `scale_policy="fallback"|"strict"|"permissive"`. The safe default retries preferred ISO
+  5455 reductions and returns the largest complete scale; strict mode raises with structured
+  requirement blockers, while permissive mode is an explicit warned opt-in to the historical
+  best-effort result. Every returned `Drawing.scale_decision` reports the requested and
+  effective scales, status, attempted scales, and any semantic placement blockers.
+
+- **Hole and hole-pattern requirements now participate in semantic completeness accounting**
+  (#1143). `lint_summary()["quality"]["completeness"]` reconciles recognition-owned bore,
+  depth/through, grouping, pattern, and location requirements to placed, suppressed, dropped,
+  missing, or explicitly unverifiable outcomes. Ambiguous declared/automatic correspondence
+  fails closed without shrinking the physical denominator. Z-normal hole locations retain ADR
+  0016's feature-level `location.location` addressability while structured physical evidence
+  distinguishes the independently required X and Y ordinates for critique.
+
+- **`lint_summary()` exposes drawing-quality evidence as separate completeness, restraint,
+  and legibility components** (#1127). The legacy `score` remains byte-compatible and is also
+  returned as the honestly named `diagnostic_score`; no composite quality verdict is invented.
+  Completeness scores recognition-owned requirements in the families with semantic outcome
+  ledgers and states that conditional scope explicitly; recognized families not yet scored are
+  listed separately. Its scalar is `audited_score`, not `score`, because a feature recognition
+  missed never became a requirement — so a part with a recogniser gap can reach 1.0, and the
+  qualifier belongs where it survives being quoted. The block also lists what its denominator
+  `excludes` and counts `unrecognised_geometry_reports` (a floor on that gap, not a measure).
+  It is not a completion gate. Legibility includes only layout and placement diagnostics, and
+  scores information-severity ones — "place what fits" drops, a leader crossing a silhouette —
+  against the warning floor, so it cannot read 1.0 while itemising output it calls unreadable.
+  A drop is identified by its `*_dropped` code suffix unless its producer records an explicit
+  `outcome_stage`, so a drop code added later counts without being registered anywhere.
+  Restraint fails closed as unavailable until measurement provenance can classify every
+  annotation.
+
+### Fixed
+
+- **Generic notes/data tables use every viable free sheet region and explain failed placement**
+  (#1145). A drawn sheet frame or zone ruler now defines the usable inset instead of becoming a
+  page-sized solid obstacle, so a measured four-row notes block can use clear lower-left A4
+  space. Tables retain the drafting preset's external text clearance from named and anonymous
+  existing annotations. Failed placement diagnostics report the measured page-space footprint,
+  attempted candidate regions, and the named obstacles or clearance bands that blocked the
+  nearest candidates without materialising the candidate Cartesian product.
+
+- **Automatic hole-table escalation is transactional per semantic requirement** (#1144). The
+  engine suppresses replaceable callouts and location dimensions only after the table fits and
+  every visible row has its required feature-owned balloon. A failed or partial balloon attempt
+  rolls the shared table back as a unit and restores the exact annotations, ordering, registry
+  identities, coverage, and issues. Precise ring, glyph, and leader-segment tests keep placement
+  clear of retained callout labels, leader shafts, and existing public-balloon components without
+  changing the existing additive `add_hole_table()`/`add_balloons()` behavior; malformed component
+  metadata—including Boolean cardinalities or coordinate payloads—falls back conservatively or fails the
+  optional replacement closed. New balloon shafts must also remain mutually crossing-free.
+  Compound, thread,
+  profile, pattern, tolerance, and fit facts the table cannot state remain active, while
+  independently replaceable X/Y locations may still move to the table. Guarded interval solving
+  has deterministic pre-lane, pre-carve, and work/state budgets; an adversarially fragmented label
+  inventory or band fails the replacement closed before excessive work or allocation. Required row-key
+  balloons take priority over optional non-certifying pattern markers, so an independently complete
+  table cannot be displaced by an auxiliary marker.
+  Hole outcomes identify the winning representation per requirement;
+  mixed table/feature evidence deliberately claims no single winner. Exceptions restore those
+  semantic markers, and automatic escalation fails closed rather than overwrite an existing
+  annotation that owns a reserved table or balloon name.
+
+- **Generated hole-callout leaders expose their exact rendered semantic text through the public
+  `Leader.label` field** (#1142). Geometry and placement are unchanged, and structured coverage
+  remains authoritative instead of requiring downstream tools to parse display text.
+
+- **Leader placement probes use analytic footprint arithmetic instead of constructing temporary
+  OCCT geometry** (#1138). Representative probe-heavy layouts are substantially faster while
+  preserving custom-style containment and placement behavior.
+
+### Changed
+
+- **Pairwise lint detail no longer multiplies the legibility score penalty** (#1147). Raw
+  severity and code counts retain their existing finding-level semantics, and every offending
+  pair remains in `issues`. The legibility component now also reports its primary issue counts
+  and scores one producer-identified annotation/failure mechanism once, with `affected_pairs`
+  exposing how many raw pair findings were aggregated. The legacy top-level `score` and
+  `diagnostic_score` remain unchanged.
+
+- **Pull-request validation now reports a stable aggregate `ci-ok` check and runs the supported
+  Python matrix on Linux** (#1134/#1136). Weekly, manual, and labelled workflows retain the
+  cross-platform escape hatch, reducing routine Actions cost without narrowing supported Python
+  coverage.
+
+## v0.4.3 — 2026-08-09
+
+**A trust and validation-throughput patch release.** Whole-part hexagonal stock now receives its
+manufacturing definition, assembly recognition no longer combines evidence across bodies, and
+performance validation is both more stable and faster to run.
+
+### Added
+
+- **Whole-part regular hexagonal stock is recognised independently of an attached boss and
+  dimensioned as `HEX … A/F` plus its axial length** (#1082). In-plane rotation preserves the
+  geometric across-flats size instead of substituting an orientation-dependent bounding-box
+  width. Automatic and declared paths share the same compiler and placement machinery, while
+  completeness lint distinguishes placed, authored-suppressed, dropped, missing, and ambiguous
+  outcomes.
+
+### Fixed
+
+- **Slot and pocket patterns cannot group members from separate assembly bodies** (#1073).
+  Recognition carries solid ownership through grouping and fails closed when ownership is
+  unavailable, preventing plausible-looking phantom patterns across disconnected parts.
+
+- **The annotation-obstacle performance guard uses deterministic work-count evidence instead of
+  a wall-clock threshold** (#1071). The regression test remains load-bearing under mutation but
+  no longer flakes because of machine load or suite ordering.
+
+### Changed
+
+- **The post-merge slow tier distributes individual tests across workers** (#1108), using
+  `pytest-xdist`'s load scheduling after local benchmark evidence showed a material speed-up
+  without weakening fixture coverage. Hosted slow validation remains the single post-merge gate.
+
+## v0.4.2 — 2026-08-09
+
+**The trustworthy manufacturing drawings release.** AP242 PMI is now accounted for from its
+source records through lowering and rendering, imported datum faces retain an auditable topology
+identity, and several drawing paths that previously omitted, invented, or stalled on defining
+geometry now fail closed or produce the required annotation.
+
+### Added
+
+- **AP242 geometric tolerances and datum features survive as source-backed manufacturing
+  requirements** (#62/#675/#1094/#1095/#1099/#1100). The importer recovers range-encoded
+  dimensional tolerances, tolerance characteristics, magnitudes, datum chains, and supported
+  all-around/all-over modifiers. Datum representation-item faces are resolved to imported
+  topology through OCCT's exact transfer relationship rather than geometric resemblance, so the
+  NIST CTC-01 datum definitions `A`, `B`, and `C` lower and render without guessing.
+
+- **PMI completeness lint reconciles every AP242 source record through extraction, projection,
+  lowering, and rendering** (#623). Unsupported or incomplete requirements remain explicit with
+  source-specific outcomes instead of allowing a partial PMI transcription to report clean.
+
+- **Hexagonal and other regular polygonal bosses are recognised and dimensioned by their
+  across-flats size** (#676). Automatic and declared paths share the same feature model, and
+  mutation tests reject incomplete, unequal, non-planar, and ambiguous face sets.
+
+- **Generated `Sheet` scripts preserve references to named source objects where correspondence
+  is proven** (#1041). The `--script` path emits declarations such as
+  `sheet.hole(features.tap)` instead of copying detected numbers; ambiguous, stale, non-shape, or
+  unused mappings fail closed.
+
+- **A curated API reference is published at
+  <https://pzfreo.github.io/draftwright/>** (#846), covering the supported entry points,
+  `Sheet` declaration surface and fluent handles, `Drawing` results, and feature constructors.
+
+### Fixed
+
+- **Recognition no longer combines faces from different solids into phantom slots or pads**
+  (#958). Each physical solid is recognised independently before the results are combined.
+
+- **`Sheet.from_part(...).envelope()` reuses the detected whole-part envelope** (#1000) instead
+  of appending a duplicate feature whose dimensions were then silently deduplicated on the page.
+
+- **Crowded Z-turned step chains recover the independently approved overall height when the
+  chain cannot be placed** (#955), and placed step lengths retain measurement identity for
+  differential audit (#1004).
+
+- **All-over GD&T leaders and complex STEP B-spline paths have bounded, export-safe DXF
+  conversion** (#1097/#1070). All-over scope now reaches the existing export-safe leader
+  representation, while spline conversion reads OCCT knots, poles, and weights through its
+  indexed API instead of expensive sequence wrappers; visible curve geometry is preserved.
+
+### Changed
+
+- **Hosted CI retains every supported Python version and both dependency/kernel generations on
+  macOS and Windows with a smaller compatibility matrix** (#1103). PRs still require lint,
+  coverage, Codecov, and the compatibility gates; the slow standards tier runs once after merge.
+
+- **The pinned wheel fixture now records mutation-gated evidence for a repeating 13-fold radial
+  profile without claiming gear semantics** (#1085). Runtime recognition remains fail-closed
+  until a separate feature contract has independently sufficient evidence.
+
+## v0.4.1 — 2026-08-07
+
+**The recognition trust release.** Automatic drawings now fail closed on unsupported inner
+profiles, prove semantic coverage for flats, slots and patterns, and retain the measurement
+identity needed to explain placed, suppressed, dropped and missing definitions. Several real
+STEP case studies also move from clean-looking but incomplete output to complete drawings or
+actionable lint.
+
+### Discouraged (supported)
+
+- **`Sheet.auto_dimensions()` and `Sheet.add_dimension()` now emit
+  `SoftDeprecationWarning`** (#1043). They remain supported with no removal planned; the
+  warning steers declared scripts toward `authored_dimensions()` plus explicit
+  `dimension(feature, parameter_id)` lines, where omission can mean suppression. Automatic
+  `build_drawing(part)` and the `Sheet.from_part(part)` on-ramp remain silent. See
+  `docs/deprecations.md` for filtering and migration guidance.
+
+### Added
+
+- **Finished drawings expose why measurements were omitted and, on covered placement paths,
+  what an annotation measures** (#996/#1002). `Drawing.suppressions()` returns the compiler's
+  omission ledger, `Drawing.measurement_keys(name)` exposes recorded measurement provenance, and
+  `draftwright.audit.diff_builds(before, after)` reports losses, substitutions and candidate
+  suppression explanations. The differential is deliberately a triage aid, not a proof where
+  renderer identity remains incomplete.
+
+- **Physical-completeness lint now follows semantic correspondence for flats, lone slots and
+  slot patterns** (#1018). Coverage is joined through recognition, IR and compiler identities
+  rather than inferred from label text or page geometry, and distinguishes authored
+  suppression, placement drops, missing definitions and unverifiable provenance.
+
+- **Full-span floored open channels are recognised and dimensioned as first-class features**
+  (#917). Automatic and declared drawings state the independent overall extent, one wall and
+  channel width, suppress the derived opposite wall, and report any missing member of that
+  defining chain.
+
+### Fixed
+
+- **Square and near-square parts retain both independent planar extents** (#997). A plain
+  square plate could previously lose both plan dimensions, while parts up to five percent off
+  square could lose one and be silently represented as square. The unsafe suppression rule is
+  removed; explicit square notation remains tracked separately in #918.
+
+- **Machined flats retain their physical stock identity from recognition through placement
+  and completeness lint** (#1013/#1015/#1034/#1036). Flats on separate parallel, coaxial or
+  slanted stock no longer collapse into one definition or borrow an opposite face from another
+  stock region. Independent callouts receive clear-margin candidates, while double-D faces on
+  one stock still form one A/F requirement.
+
+- **Dense callout and step-detail cases keep their complete defining set** (#915). Hole
+  callout batches are reconciled as a queue, detail views fit their actual aspect, step levels
+  retain the supporting-face correspondence needed for truthful crops, and a detail redraws
+  only rungs omitted from its parent view.
+
+- **Pocket leaders start on the physical opening rim** (#916), removing the spurious inner
+  silhouette crossing while preserving the solved label corridor and approved dimensions on
+  X-, Y- and Z-depth pockets.
+
+- **Wedge-mounted raised pads survive recognition** (#909). Lower ledges that touch only in
+  plan no longer make the recogniser discard a valid upper pad; true staircase tiers remain
+  excluded.
+
+- **Annotation-dense detail placement no longer stalls the build** (#1065). Detail views
+  avoid every decomposed annotation shaft, witness, label and item footprint; CTC-02
+  supplies 571 such boxes. The free-rectangle search previously treated that set as tiny
+  and admitted roughly 34 billion candidate rectangles, exhausting the slow tier's
+  ten-minute test budget before export. A compressed-coordinate sweep preserves the same
+  deterministic winning rectangle while completing the real search in seconds.
+
+- **Through double-D bores are recognised and called out as first-class hole profiles**
+  (#1061). Automatic and declared builds preserve the parent-circle diameter, A/F size,
+  orientation and depth, and render one compound `DOUBLE-D ... A/F` bore callout. Physical
+  critique now accepts that supported inner profile while retaining a separate warning for
+  13 evenly spaced common-circle arcs on the unresolved outer boundary; it does not infer a
+  repeating full profile or a gear standard.
+
+- **Unsupported internal profiles no longer pass physical-completeness lint silently**
+  (#1058). Inner loops on principal boundary faces are checked against the current
+  circular-hole, rectangular-opening, true-obround, and proven through-double-D vocabulary.
+  Profiles outside it now produce
+  `unrecognised_defining_geometry` and reduce the lint score instead of allowing an
+  envelope-only drawing to score 1.0.
+
+### Changed
+
+- **Crowded prismatic step dimensions recover into an enlarged detail view by default**
+  (#909). The automatic `build_drawing` / `make_drawing` / CLI paths and generated
+  `Sheet` scripts now preserve the omitted dimensions instead of returning
+  `step_dim_dropped` unless the caller remembered `detail_view=True`. Pass
+  `detail_view=False` to retain the parent-view-only behavior and its lint warning.
+
+## v0.4.0 — 2026-08-02
+
+**The compat-exit release.** Every surface whose documented removal target was 0.4.0 is gone,
+and every deprecation that remains now says when it goes — in the message the caller sees, not
+only in a release note. ADR 0005 §4's rule is that an alias carries a tracking issue *and* a
+removal date, because "a facade with no exit date is a failure mode"; this release is that rule
+being kept rather than restated. `docs/deprecations.md` is the new index, and
+`tests/test_deprecation_dates.py` fails any deprecation that names no removal.
+
+Also lands ADR 0016's declared-dimensioning work (#867): a `dimension(feature, role)` is
+referential — it names a measurement and carries no number — and an authored set means omission
+is suppression, enforced at the compiled-plan boundary rather than by a flag renderers check.
+
+**Read this before upgrading:** two removals break *without a release that warned you*, because
+their deprecation never appeared in a shipped version. See *Removed (breaking)* below and
+`docs/deprecations.md`.
+
+### Removed (breaking)
+
+Compat surfaces whose documented removal target was 0.4.0 (#720, ADR 0005 §4 — every alias
+carries a tracking issue *and* a removal date, or the facade is permanent by accident):
+
+- **The seven `Drawing` compat aliases** — `_named` / `_anno_view` / `_pinned` /
+  `_build_issues`, and `_pattern_callouts` / `_patterned_holes` / `_dropped_callout_diams`.
+  These were private, and the public reads have existed since #699: use `dwg.registry`
+  (`in reg`, `names()`, `issues`, `pinned_names()`) and `dwg.coverage`, or the `Drawing` verbs
+  `annotations()` / `iter_annotations()` / `get_annotation()` / `view_of()`.
+  **Note the asymmetry.** A *read* — `dwg._named` — now raises `AttributeError`. A *write* —
+  `dwg._pinned = {...}` — does **not**: these were properties with setters, and `Drawing` has
+  no `__slots__`, so assignment now quietly creates an unrelated instance attribute that no
+  longer reaches the registry or coverage owner. Writes therefore fail silently rather than
+  loudly. Grep for `\._(named|anno_view|pinned|build_issues|pattern_callouts|patterned_holes|dropped_callout_diams)\b`
+  before upgrading; mutate through `registry.add()` / `pin()` / `record_issue()` /
+  `restore_issues()` and the `CoverageState` methods instead.
+- **The `draftwright.sheet_dsl` module** — an import alias for `draftwright.sheet` since the
+  #640 rename. Import `Sheet` from `draftwright` or `draftwright.sheet`.
+- **`generate_script`**, including its `draftwright.__all__` entry. It has raised since #940
+  retired the imperative emitter; it is now simply absent, so the failure is an `ImportError`
+  at the top of a script rather than a `RuntimeError` part-way through one. Use
+  `--script` / `emit_sheet_script`.
+- **The bespoke `--style imperative` error message.** `--style` itself is unchanged and still
+  accepts its sole value `sheet`; `imperative` is now an unrecognised value like any typo,
+  rather than one carrying its own explanation of the #940 retirement.
+
+- **Bare dimension-role spellings** — `sheet.dimension(f, "width")`. Use the parameter id,
+  `"width.length"`; `dimension_ids()` on a handle lists the valid ones. The bare role is the
+  *family* spelling: it selects every parameter carrying it, which is how
+  `dimension(step, "step")` quietly declared two measurements. In an authored set, where
+  omission means suppression, silently declaring an extra one is the mirror image of the rule
+  — so it now raises rather than resolving. (A *discriminated* bare role, used with `axis=`,
+  is unaffected: that is how variants like `grid_pitch.length.row` are addressed, and it never
+  warned.)
+- **The `Sheet.dimension(kind=…, value=…)` call shape** — use `Sheet.measured_dimension(...)`.
+  `dimension` is now solely the ADR 0016 referential verb: it names a feature and a parameter
+  id and reads the value off the geometry.
+
+Those last two break **without a release that warns you**. They were deprecated after v0.3.9
+and removed in 0.4.0, so the `DeprecationWarning` never appeared in a released version —
+upgrading from v0.3.9 or earlier goes straight from working to a raise. That is deliberate
+(ADR 0016); this entry and `docs/deprecations.md` are the only notice you get, so both
+failures name their replacement rather than raising about argument counts.
+
+### Deprecated
+
+- **The legacy `Drawing.export` shapes now emit `DeprecationWarning`** (#987) — they were
+  announced as deprecated in v0.3.1 and then said nothing at runtime for four minor releases,
+  which would have made their 0.5.0 removal a silent break. Three cases warn:
+  - `export()` with `formats=` **omitted or `None`** — the legacy default, which writes
+    SVG + DXF and returns an `(svg, dxf)` **tuple** rather than the `{format: path}` dict.
+  - the **`svg=` / `dxf=` booleans**. The suggested replacement names the formats *that call*
+    selected, so following it cannot change what gets written.
+  - passing a boolean **alongside** `formats=`, where it is silently ignored — `formats` wins,
+    and now says so.
+
+  **If you promote `DeprecationWarning` to an error, previously-passing exports will now
+  raise.** Migration: `export(out, formats=("svg", "dxf"))` and read the dict. `make_drawing`
+  is unaffected — it returns the same tuple and does not warn.
+
+### Added
+
+- **`sheet.add_dimension(feature, role)` — ask the planner to carry one more
+  measurement** (ADR 0016 / #872). Referential: it names a feature and a role and
+  carries **no number**, so the value still comes from the geometry and a size lives in
+  exactly one place. It changes *selection*, not derivation — a request can never
+  introduce a number the part does not have. Requesting something the planner already
+  emits is a deliberate no-op, so a script can ask without first knowing the rule set's
+  mind. `sheet.auto_dimensions()` states the source explicitly (optional in this
+  release; #874 makes it mandatory).
+- **Suppressing a dimension marks it; it no longer leaks into the callout** (ADR 0016 /
+  #875). `PlannedDimension.suppressed` was honoured at thirteen render sites but not by
+  the compound hole-callout path, so a suppressed counterbore still printed. The group
+  keeps its engineering data either way — what changes is whether a value reaches the
+  page. Suppressing the bore ⌀ while a counterbore, spotface or countersink segment
+  remains now **raises** and names the orphan: `⌀20 THRU ⌴ ⌀32 ↓ 1.5` has no reading with
+  its leading term removed, and silently dropping the segments would discard authored
+  intent while silently restoring the head would make the drawing say something the
+  script does not.
+- **Addressable dimension identity** (#869/#870/#871): every planned measurement now has
+  a stable `DimensionId(feature, parameter)`, and correlated measurements that must be
+  named as one — a grid pattern's row and column pitch, a step ladder — group into an
+  `AddressableDimension`. This is what a `dimension(...)` line will name, and what
+  suppression and provenance key on.
+
+### Changed
+
+- **`Sheet.dimension(kind=…, value=…)` is now `Sheet.measured_dimension(...)`**
+  (ADR 0016 / #873), and `model.declare.authored_dimension` is
+  `model.declare.measured_dimension` to match. This **reserves** the name `dimension` for
+  the referential verb ADR 0016 defines — name a feature and a role, carry no number, let
+  the engine read the value off the geometry — which `Drawing.dimension` already is and
+  `Sheet.dimension` will become. The verb that carries an explicit number needed a name
+  saying so first. The old spelling was a transitional overload during development and is
+  **removed in this same release** (#720) rather than shipping — `Sheet.dimension` is solely
+  the referential verb, and the old keyword call raises with `measured_dimension` named. See
+  *Removed (breaking)* above. Generated AP242 scripts emit `measured_dimension` and so arrive
+  un-deprecated.
+- **`Sheet` handles address features by identity, not position** (#908/#910/#912). A
+  handle, tolerance, GD&T origin, section request or dimension intent now follows *its*
+  feature through a `features` reorder instead of naming whatever took the slot.
+  Reordering with `reverse()`/`sort()` preserves every reference; deleting or replacing
+  a referenced feature raises rather than silently retargeting a neighbour. `of()` now
+  accepts a handle, and negative indices resolve uniformly across `of()`,
+  `add_dimension()` and the GD&T verbs.
+
+### Fixed
+
+- **`sheet.envelope()` measured the file, not the part** (#977). An AP242 STEP import is a
+  compound of the solid *plus its PMI presentation geometry* — annotation planes, leader
+  curves — and the declared verb measured all of it. On the NIST CTC-01 fixture it declared
+  1170 × 650 where the part is 800 × 450: an envelope 370 mm too wide, silently, in anything
+  declared by hand. `step_level()` had the same exposure, offsetting every step position by
+  the annotation overhang. Both now measure the solid body, sharing the helper the engine
+  already used for exactly this (`_analyse` and `Sheet.model`).
+- **A declared envelope's frame origin sat a half-height below a detected one** (#977). It
+  used `bbox_min.Z` where the detector uses the bbox centre. **Dimension output is unchanged** —
+  envelope sizes derive from `bbox_min`/`bbox_max`, and the overall height is compiled from the
+  model bbox — but `frame.origin` is the generic feature *site*, so **a GD&T control frame,
+  surface finish or note targeting a hand-declared `sheet.envelope()` was anchored at the
+  bottom face rather than the centre, and its leader now moves to the centre**. That is a
+  visible change on drawings carrying such an annotation; drawings without one are unaffected.
+  Scripts generated by `--script` were never affected either way: they bake the detected frame
+  explicitly rather than calling the verb.
+
+
+- **Front-view dimensions join the placement solve instead of committing to the strip**
+  (#894): they previously took strip space before the global solve ran, so a
+  higher-ranked dimension could find its space already gone. Priority only ranks
+  candidates that are *in* the solve.
+- **A hole callout reads the feature's own `through` fact** (#868) rather than inferring
+  it from whether a depth parameter happens to be present — so a suppressed depth can no
+  longer make a blind hole print `THRU`.
+- **The balloon ring clears the view it annotates after a hole-table escalation**
+  (#901/#903).
+
+## v0.3.9 — 2026-07-26
+
+### Added
+
+- **Repeated recesses are dimensioned as one pattern, not N competing callouts**
+  (#841): identical blind pockets and identical through-slots now collapse to a single
+  grouped `5× 7.9 × 13.6 × 19 DEEP` / `4× SLOT 8 × 30` leader plus `(n-1)× pitch`
+  dim(s). Previously each member competed for its own size dims and some silently
+  dropped for want of strip room — five declared slots rendered only three length dims.
+  Both new kinds (`PocketPatternFeature`, `SlotPatternFeature`) round-trip every
+  surface: declare (`sheet.pocket_pattern(...)` / `sheet.slot_pattern(...)`),
+  recognition from a real solid (coplanar, same-facing members only), the emitted
+  Sheet script, and the editable surface (`dwg.callout(feature)` + `finalize()`).
+- **`Sheet.section()` and `Sheet.detail()` — ask for the view you need** (#841/#847):
+  a section A–A auto-fires only for a Z-axis hole with a counterbore, spotface, or
+  blind bottom, so a blind pocket's floor and depth stayed hidden-line-only with no
+  supported way to request a cut. `sheet.section(feature)` / `section(at=y)` /
+  bare `section()` force one (the cut plane is validated to lie strictly inside the
+  part), and `sheet.detail()` exposes the existing `detail_view=True` opt-in as a
+  chainable verb.
+- **External threads are a first-class declared aspect** (#859): the turned analog of
+  the #764 internal thread — `sheet.step(shaft).thread("M3x0.5")`,
+  `sheet.diameter(...).thread(...)`, and `sheet.boss(...).thread(...)` append the
+  spec to the ⌀ callout (`ø3 M3x0.5`) instead of needing a free-text `.note()`. It
+  composes with `.finish()` for Ra-on-thread, keys the callout bucket on `(⌀, thread)`
+  so a threaded and a plain ø6 stay distinct, and now round-trips through the Sheet
+  emitter — including pattern-member holes, closing the symmetric latent #764 gap.
+  Declaration-only: a plain cylinder is geometrically indistinguishable from a
+  threaded one, so there is no recogniser. *(Known limitation: the incremental
+  `callout`/`finalize(only=…)` paths still dedup thread-blind — #863.)*
+- **Blind obround pocket recognition, including imported STEP** (#837): the blind
+  counterpart of the #816 through-slot work. A stubby floored obround has side walls
+  too short to pair, so it is recovered from its semicircular end caps — clustering
+  faces by axis proximity so the quarter-cylinder split a STEP importer commonly
+  produces recognises as well as build123d's half-cylinder. A sealed internal void
+  (capped at both ends) is correctly *not* a pocket. A real tuner-jig STEP with five
+  blind pockets ships as a fixture.
+- **`sheet.slot(...).note(...)` / `sheet.pocket(...).note(...)`** (#841/#845): the
+  slot/pocket handle now anchors a note to its own feature, with no `ref=` needed
+  (an explicit `ref` still forwards, as before).
+
+### Changed
+
+- **An anchored note, GD&T frame, or finish relaxes its side instead of silently
+  vanishing** (#841/#855): an annotation declared with an explicit `view=`/`side=`
+  used to drop with only a warning when that strip was full, while the export still
+  reported success. The requested side is now a *preference* — when the strip has no
+  room the placer tries the opposite side, then the two perpendicular sides, taking
+  the first with room (via the same bounds and title-block checks, so it never
+  overshoots). A relaxed placement records an INFO `gdt_side_relaxed` issue naming
+  requested versus actual. Only when no strip fits does it drop.
+
+### Internal
+
+- **The state-bus endgame is closed to its permanent seam** (#830/#840): the
+  solve trace is filled into `BuildState` at the one construction site, the redundant
+  part-model re-attach is gone, and the detail view is now *transactional* — geometry
+  is projected at the detail scale and committed only if dimensions land, so
+  `_drop_view_coordinates` has zero engine callers. With those removed, the #817
+  decision-D method-call exemption narrows from a blanket pass to a named
+  `_LAYOUT_SEAM` allowlist (`_add_view`, `_set_view_coordinates`), accepted as the
+  permanent interactive-layout seam; every other private-method call on the drawing
+  is now flagged.
+- **A draftwright logo set** (#843) — drafting-idiom mark, all text outlined so the
+  SVGs render font-independently; the README now leads with the lockup.
+- **A worked multi-feature object-reference example** (#853) in
+  `docs/multi-feature-object-reference-workflow.md`.
+
+## v0.3.8 — 2026-07-23
+
+### Added
+
+- **`Drawing.note()` — a public free-text note verb** (#817/#820): place a free-form
+  `Note` at a page point without reaching for the low-level placement primitive —
+  `dwg.note("SEE NOTE 1", at=(x, y))`. This is the safe door for free text that the
+  privatisation below needs.
+- **Obround (racetrack) through-slot recognition** (#816): a part whose only features
+  are obround through-slots is now recognised, dimensioned, and referenceable for GD&T.
+  Previously such slots were silently undetected — and lint falsely reported the drawing
+  complete. Stubby obrounds, whose straight side walls are too short to pair, are recovered
+  from their two semicircular end caps (a round hole is never mistaken for a slot end).
+
+### Changed
+
+- **The generated Sheet script is now self-describing** (#833): `--script --style sheet`
+  emits a feature-census header, section sub-headers with repeat tallies (`8× R50`), and a
+  plain-language comment on each verb line (`# ⌀8 THRU ×4`, `# slot 40 × 120`, `# R50`).
+  All additions are comment-only, so re-run fidelity is unchanged.
+
+### Deprecated
+
+- **The low-level `Drawing` placement API is now private — the public surface is
+  safe-by-default** (#817): `add`, `place_dim`, and the view-coordinate / build-state
+  plumbing (`set_view_coordinates`, `drop_view_coordinates`, `attach_part_model`,
+  `attach_solve_trace`, `add_view`, `clear_annotations`) are deprecated in favour of the
+  feature-backed verbs (`callout` / `dimension` / `locate` / `note` / …), which route
+  through the layout solve. Each keeps a `@deprecated` (PEP 702) shim for one release —
+  so type-checkers and IDEs flag call sites, not just a runtime warning — and is slated
+  for removal in ~0.5.0.
+
+### Internal
+
+- **`AGENTS.md` — an AI-agent usage guide** (#818) with an executable check, steering
+  agents toward the safe declarative surface rather than the low-level primitives.
+- **Coverage reporting is governed and a Codecov badge added** (#825/#832).
+
+## v0.3.7 — 2026-07-21
+
+### Fixed
+
+- **Turned ⌀ leaders point at the feature, not a corner or across the body**
+  (#794/#798): the diameter callouts on a turned part now centre on the step's
+  mid-length (a boss keeps its frame origin; a ⌀ shared by disjoint steps centres
+  on the longest run), so the arrow lands on the middle of the feature's edge
+  rather than a step corner. A leader that would otherwise cut diagonally through
+  the body — a nested or end feature such as a ⌀6 boss stub — is re-routed out to
+  the nearest clear margin (axis-aware: X-turned → left/right, Z-turned → top/
+  bottom), pointing straight at the feature instead of clipping the silhouette. A
+  re-route that can't be placed clear-and-safe is restored and flagged, and a
+  pinned leader is never moved.
+
+### Added
+
+- **`leader_crosses_silhouette` lint notice** (#796): an info-level structural
+  notice when a leader shaft passes *through* the projected part body (as opposed
+  to a hole callout legitimately exiting an internal hole), so the gap is
+  observable even where routing cannot clear it.
+
+## v0.3.6 — 2026-07-21
+
+### Fixed
+
+- **`Sheet.step(boss)` on a prismatic part fails loudly instead of silently
+  dropping the height** (#631): declaring a turned `.step()` at a cylinder that is
+  really an external boss set `orientation="z"`, which made `render_height_ladder`
+  suppress the overall-height dim on the assumption the OD/step chain conveyed it —
+  but on a prismatic box that chain never renders, so the height vanished and
+  nothing replaced it (11 → 10 annotations). The build now detects a `StepFeature`
+  coincident with a `BossFeature` (same axis, origin, diameter) — a combination
+  that can never be legitimate — and raises a clear `ValueError` pointing to
+  `.boss()` (which renders its own ø and height per #632). Legitimate turned shafts
+  have steps but no coincident boss, so they are unaffected.
+- **`detail_view=True` no longer a silent no-op when the detail can't be placed**
+  (#630): a part whose lint recommends a detail view (`step_dim_dropped`) but whose
+  crowded band is full-width (e.g. a shelled cover's stacked face levels) can't be
+  enlarged legibly *and* still fit alongside the main views — so `detail_view=True`
+  produced a drawing byte-identical to `False`, silently ignoring the opt-in. It
+  now records a `detail_unplaceable` warning naming the reason and an actionable
+  remedy (dimension manually / move onto its own sheet), so the opt-in is
+  observable. Parts whose detail *does* fit are unchanged.
+- **Imperative `--script` round-trips the title-block fields + sheet furniture**
+  (#775): the imperative flavour reproduced only title/number/tolerance/drawn_by/
+  scale/page, dropping the `material`/`date`/`revision`/`company` fields (#766) and
+  the `frame`/`zones`/`projection` furniture (#767/#768/#769) — so a regenerated
+  imperative script didn't match the direct CLI. They now ride the emitted script's
+  cog config block and its `build_drawing(...)` call (the Sheet flavour already
+  round-tripped them), and the CLI forwards them to `generate_script`.
+
+### Internal
+
+- **Fast-tier dense-sheet canary for the #733/#734 strip-pressure regression**
+  (#737): a synthetic contended-front-strip fixture that drops a step-height
+  principal dim on the pre-#734 commit and stays clean on `main` — closing the
+  fast-tier gap the #733 regression fell through (only the CI-only slow CTC
+  fixtures exercised real strip pressure before).
+- **Layout-hypothesis fuzz tier is now a reproducible PR gate** (#692):
+  `derandomize=True` makes every run generate the same fixed example set (with
+  `print_blob` for paste-able reproduction), so a failure recurs deterministically
+  instead of surfacing on a seed-dependent minority of runs; exploratory fuzzing
+  moves behind `DRAFTWRIGHT_FUZZ_EXPLORE=1`. The two `test_make_drawing.py` xdist
+  observations (#669) were investigated and found non-reproducible — both tests are
+  non-flaky by construction/measurement on current `main`.
+
+## v0.3.5 — 2026-07-20
+
+### Fixed
+
+- **Generated `Sheet` script round-trips `--frame` / `--zones` / `--projection`**:
+  the sheet-furniture flags added in #767/#768/#769 were not emitted into the
+  generated `Sheet(...)` constructor (nor forwarded by the CLI's `--script` path),
+  so a regenerated script dropped the border / zone ruler / projection symbol —
+  the script no longer matched the direct CLI drawing. The emitter now emits
+  `frame=True` / `zones=True` / `projection=…` when set (a plain drawing keeps a
+  clean constructor), and the CLI forwards them to `generate_sheet_script`.
+
+### Added
+
+- **ISO 5457 zone-grid border ruler** (#768): `build_drawing(zones=True)` /
+  `Sheet(zones=True)` / `--zones` draws the grid reference system — numbers 1..
+  along the top/bottom edges, letters A.. (skipping I/O) down the sides — in the
+  band between the frame and the page edge, with the standard division count per
+  A-series page (~50 mm zones). Implies a frame (the ticks sit on it). Off by
+  default → byte-identical. The border ticks are lint-exempt like the frame; the
+  labels are exempt only from the page-bounds check (they legitimately sit outside
+  the drawable), still covered by overlap lint.
+- **Projection-method symbol** (#769): `build_drawing(projection="third"|"first")` /
+  `Sheet(projection=…)` / `--projection` draws the ISO 5456-2 third-/first-angle
+  glyph in the reserved title-block band (above the title block). Uses the new
+  `ProjectionSymbol` primitive from build123d-drafting-helpers **0.14.1** (the pin
+  is bumped). Off by default → byte-identical. Unlike the page-spanning frame, the
+  small well-placed glyph is *not* lint-exempt — lint covers it so a future
+  mispositioning is caught.
+- **Opt-in drawn sheet frame / border** (#767): `build_drawing(frame=True)` /
+  `Sheet(frame=True)` / `--frame` draws a border rectangle at the page margin. It's
+  the *content boundary* (Option B), not a rectangle over the drawing: turning it on
+  raises the effective content margin, and because `compose._layout_geometry` is the
+  single authority shared by scale/page selection **and** placement, the reservation
+  flows through `choose_scale` — so a framed drawing may pick a smaller scale / larger
+  page (ADR 0004), with content clearing the border. Default `frame=False` is
+  byte-identical (guarded by the golden + layout-cleanliness suites). The page-spanning
+  border carries an `is_sheet_frame` rider so lint skips it.
+
+## v0.3.4 — 2026-07-20
+
+### Changed
+
+- **Imperative `--script` output defaults to PDF and the modern export form**
+  (#709). **Breaking** for regenerated imperative scripts: without `-f`, a
+  regenerated script now writes `<stem>.pdf` (was SVG+DXF via the deprecated
+  legacy tuple), and the script-visible variables changed from
+  `svg_path, dxf_path` to a `paths` dict printed in the requested-format order.
+  This aligns the imperative flavour with the Sheet flavour and the direct CLI,
+  which have defaulted to PDF since #702/#288. Migration: regenerate with
+  `--format svg,dxf` to keep the old outputs. Previously-generated scripts are
+  unaffected (they carry their own export lines).
+- **`finalize()` drains in the auto-pass's order** (#699 slice b): both build
+  paths now execute the orchestrator's one canonical `_PASS_SEQUENCE` (via the
+  shared `run_stages`), removing the hand-mirrored second orchestration in
+  `Drawing._drain_intents`. For the deferred/finalize path this reorders three
+  things toward auto-pass parity: the turned diameter/step-length set-solves now
+  place *before* the corridor drain (matching the auto-pass's obstacle
+  visibility); slots register after them; and the register-only height-ladder /
+  step-position stages register *after* locations — observable too, since
+  registration order decides corridor key-creation (= drain) order and
+  same-priority tie-breaks (Codex review; the new order is exactly the
+  auto-pass's). Recompose output may shift accordingly; the auto-pass order is
+  unchanged.
+- **`Sheet.export` speaks the modern export API** (#702): it now takes
+  `formats=` (any of svg/dxf/pdf/png, default **PDF** — matching the CLI
+  default) and `dpi=`, and returns the `{format: path}` dict. Previously it
+  rode the deprecated legacy path, could only write SVG+DXF — the flagship
+  facade couldn't produce a PDF — and returned the old `(svg_path, dxf_path)`
+  tuple. **Breaking** for callers unpacking that tuple: use
+  `sheet.export(stem, formats=("svg", "dxf"))` and read the dict.
+
+### Added
+
+- **Solve-trace / explain mode** (#736, from the #733 post-mortem): opt-in
+  observability for strip placement. `build_drawing(trace=…)` or
+  `DRAFTWRIGHT_TRACE=<path-or-dir>` writes ONE JSON file per build (schema v2)
+  with two record types: `solves` — every corridor solve's key, strip bounds,
+  candidate set, the obstacles that carved the strip (with owning annotation
+  names), the free segments, and each candidate's outcome (placed /
+  dropped-with-reason / deduped / promoted / deferred-to-post-drain) — and
+  `pass_events` — the standalone strip passes plus the *immediate* placers
+  (post-drain machined-feature leader callouts, the turned diameter row/column
+  and step-length set-solves), each with per-item placed/dropped outcomes. So
+  "why did X drop" is one `jq` query (`.solves[].outcomes[]` /
+  `.pass_events[].items[]`). A `placement_unsatisfiable` strip-full drop now
+  also names the top occupants in its lint message (`…strip full; occupied
+  by: …`). Recording-only: an unwritable path logs a warning (never aborts a
+  build), writes are atomic, and the recorder joins `finalize()`'s #647
+  rollback. Default off; zero output change, nil cost when off.
+
+- **`draftwright.model.authored_dimension`** (#704): the IR constructor behind
+  `Sheet.dimension()`, extracted so `build_drawing(model=…)` callers can author
+  a pre-measured drafting dimension without the `Sheet` façade.
+
+### Added
+
+- **Object-reference script workflow: docs + an inline tip** (#770/#771): a
+  generated `Sheet` script from a **live-source** input (a `module:attr` /
+  `file.py:attr` object spec, or a build123d object) now carries an inline tip
+  showing how to swap each detected-numbers line for a reference to your object
+  (ADR 0011 declare — the size is read off the object). A STEP-sourced script is
+  unchanged (byte-stable — no object to reference). The README documents the full
+  from-a-part-to-an-object-referenced-script workflow.
+- **Knurl callout verb** (#765): `sheet.diameter(shaft).knurl("0.8")` →
+  `KNURL 0.8 STRAIGHT` (or `.knurl("0.8", "DIAMOND")`). Named sugar over
+  `.note()` with canonical formatting — a knurl is a text callout on a leader,
+  not modelled geometry, so it flows through the existing note path (no new
+  IR/render).
+- **Standing ISO 7200 title-block fields** (#766): `material`, `date`, `revision`,
+  and `company` (legal owner) are now first-class — settable on `build_drawing` /
+  `make_drawing`, the `Sheet(...)` constructor, and the CLI (`--material` /
+  `--date` / `--revision` / `--company`), and emitted into the generated `Sheet`
+  script when non-default so a re-run reproduces them. Defaults preserve the prior
+  output (revision `A`, the rest blank). The `revision`/`legal_owner` cells the
+  renderer previously hardcoded now flow from these. (The imperative `--script`
+  flavour's title-block round-trip of these fields is a follow-up.)
+- **First-class thread/tap callout on a hole** (#764): ``sheet.hole(...).thread("M3x0.5")``
+  (and ``declare.hole(..., thread=)``) folds a tap/thread spec onto the hole's existing
+  compound leader (e.g. ``ø2.5 THRU M3x0.5``) — a structured aspect that round-trips, so
+  ``.thread(...).finish(...)`` yields Ra-on-thread. A declaration-only aspect (ADR 0011
+  side-layer: threads are cosmetic, not modelled geometry — no recogniser). The layout
+  width estimator (``_est_planned_bore_callout_width``) now accounts for the thread text so
+  the wider callout is reserved room and not dropped (the #261 estimator/render agreement).
+
+### Fixed
+
+- **Generated imperative script reconstructs side-drilled hole locations**
+  (#426/#133): an X/Y-axis (side-drilled) bore gets `dim_loc_*` position dims in
+  the direct build, but the `--script` emitter emitted a gap comment instead of a
+  `locate()` for non-Z holes (because `locate()`/`render_locations` are Z-plan
+  only, #133). The emitter now emits `dwg.locate(f)` for holes on any axis;
+  `finalize()` routes by the feature's axis — Z-plan holes drain through
+  `render_locations` as before, side-drilled bores through the whole-model
+  `_locate_off_axis_holes` pass (registered at the auto-pass's own
+  `off_axis_across`/`off_axis_along` stages, placed at the shared drain). The
+  Z-only `locate()` live contract is untouched (side-drilled locates never reach
+  it — they route to the off-axis stages). Closes the last direct-vs-generated
+  script parity gap tracked by #426.
+- **Generated imperative script reconstructs rotational furniture** (#426/#424):
+  a turned/cylindrical part's `dwg.model()` carries a `RotationalFeature`, but the
+  `--script` emitter parked it in the gap-comment list, so an executed
+  reconstruction of e.g. a plain cylinder was missing `dim_od` and the
+  `centerline_front`/`centerline_side` marks the direct build draws. A new
+  `Drawing.rotational(feature)` add-verb records the intent, and `finalize()`
+  drains it through the shared whole-model `render_rotational` at the auto-pass's
+  own `"rotational"` stage — byte-identical to the direct build (no `only=` subset,
+  fixed literal names, so no naming-seam or `⊇`-vs-`==` gap). The emitter now emits
+  `dwg.rotational(f)` for rotational features. Closes the rotational half of the
+  direct-vs-generated-script convergence.
+- **Emitted `Sheet` script now faithfully reproduces the direct CLI drawing**
+  (#707): the divergence reported against 0.3.3 on the Maquetto GRM-03 part was
+  closed by #709 (script `--format` forwarding) + #661 (finalize detail drain).
+  Locked with a vendored-fixture regression that asserts the full invariant the
+  issue named — same views, annotation inventory, page, scale **and** lint
+  between `build_drawing(...)` and the executed emitted `Sheet` — not just the
+  annotation signature the synthetic parity cases check.
+- **`_largest_empty_rect` no longer blows up on crowded detail views** (#661):
+  the largest-empty-rectangle placer enumerated every candidate rectangle
+  (O(N⁴) in the coordinate-cut count) — fine for the iso view's handful of
+  obstacles, but the detail-view placer feeds it *every* placed-annotation
+  footprint, so a two-detail turned part reached ~70 obstacles and spent ~13 s
+  in a single call (multi-minute on the slower Windows CI runners, which tipped
+  some jobs over their wall-clock budget). It now skips candidates that provably
+  can't beat the best found so far (a `bisect` past every too-narrow pair, an
+  early `break` once even the widest is too small): the same exact result and
+  tie-break — verified identical to the naïve search across 3 000 random cases —
+  at ~85× the speed on the pathological input (12.6 s → 0.15 s).
+- **Detail-view height demotion is pin- and provenance-safe** (#661, user
+  review): when a crowded prismatic detail needs the overall-height dim's room it
+  is demoted and the detail retried, but (a) the canonical `dim_height` name was
+  handed to that retry unconditionally — a *pinned* or user-replaced height could
+  be removed; it now passes the same pin/label guards as the generalised
+  `dim_length{n}` path (a pin, ADR 0012, is never demoted), and (b) if the retry
+  also failed, the restore re-added only object/name/view, orphaning the dim from
+  its envelope feature (lost `annotations_of`/`drop`/re-discovery); it now
+  snapshots and restores the feature provenance and pin too.
+- **`finalize()` resolves queued detail requests — detail views now exist on the
+  edit path** (#661): the finalize drain was missing the auto pass's
+  `detail_request`/`details` stages, so a crowded turned head's queued
+  `DetailRequest` died with the per-run context and a prismatic
+  `detail_view=True` build never even queued one — generated scripts silently
+  lost every DETAIL A/B view. `build_drawing(detail_view=…)` is now persisted on
+  the build state so the drain gates the prismatic request exactly as the auto
+  pass does; the details stage re-projects the iso at sheet scale for the
+  free-rectangle search and refits it after (mirroring the auto pass's ordering,
+  where details place before the iso is fitted); and the overall-height demotion
+  retry now finds the height dim by its envelope attribution, not only the
+  auto-pass `dim_height` name. Three script-parity characterisation tests
+  (single detail, two details, turned head) flip from xfail to passing.
+- **Y-turned stepped shafts no longer crash** (#661): `render_step_lengths`
+  projected every step span into the front view, where a Y-axis span is end-on
+  (a point) — `Dimension` then raised "start and end points must be different"
+  on BOTH the direct build and a generated script's finalize. Y-axis steps are
+  now skipped (no step render pipeline consumes Y, #731 — mirroring
+  `render_diameters`' x/z bucketing), the emitter flags the gap as a comment
+  instead of emitting an unreplayable `dimension()` verb, and a hand-written
+  Y step-length intent live-replays (surfacing the verb's own behaviour) rather
+  than vanishing into the routed no-op. The Y-axis script-parity
+  characterisation test flips from xfail to passing.
+- **Boss heights are modeled, rendered, and coverage-checked** (#632): detected
+  and object-declared cylindrical bosses now carry their axial extent through
+  `BossFeature` and the dimension planner. Prismatic bosses receive a linear
+  height dimension in a profile view in addition to their end-on diameter
+  leader. `Drawing.lint()` reconciles that modeled height against the live
+  feature-owned `Dimension`, reporting `boss_height_missing` if it is removed
+  or otherwise absent (demoted to `info` on assembly drawings). This closes the
+  declarative false-negative where a boss's diameter and the overall envelope
+  were documented but its own independent height was not.
+
+- **`--format` now reaches `--script` output** (#709, from the #702 adversarial
+  review): `--script -f svg` used to silently emit a PDF-producing script — the
+  CLI parsed the flag but never forwarded it. Both emitters now thread it
+  through: the Sheet script emits `sheet.export(stem, formats=(…,))` when
+  non-default (bare call for the PDF default), the imperative template embeds
+  the requested tuple and prints in the **requested** order, and
+  `generate_script` / `generate_sheet_script` grow a `formats=` parameter. The
+  dormant script-parity characterisation module
+  (`test_script_detail_parity.py`) is also un-skipped so its strict xfails give
+  the remaining gaps CI signal.
+
+- **Authored fillet, flat, groove, pocket, plate and slot tolerances now render**
+  (#725/#726/#727/#728/#729/#730 / #698): the four leader-callout kinds join
+  `_CONVENTION` and their renderers (`render_fillets`/`render_flats`/
+  `render_grooves`/`render_pockets`) — and the plate-thickness and slot
+  width/length linear passes
+  (`render_plates`/`render_slots`, explicit `"linear"` entries) — consume the
+  planner's `DimensionGroup`s, binding each
+  planned dim explicitly by `(role, kind)` — so a `decorations`-authored ± /
+  limit tolerance reaches the placed callout (`R8 ±0.1`, `17 ±0.2 A/F`,
+  `4 ±0.1 WIDE × ø16 ±0.5`, `18 ±0.2 × 30 ±0.2 × 5 ±0.2 DEEP`, a `10 ±0.1`
+  thickness dim, an `8 ±0.1` slot width; on a
+  multi-value label each suffix rides its own number). Previously all six
+  passes formatted raw feature fields and silently dropped it — the latent #629
+  bug class the ADR 0015 bypass list documents. The fillet `n× R` and flat
+  double-D/hex collapses stay render-side (first-authored tolerance wins, the
+  `render_diameters` precedent); a pocket's three values — and a slot's
+  width + length — share one authored
+  tolerance (all kind `"length"`). The slot's model-derived datum position dim
+  and its corridor/immediate placement mechanics are untouched. Untolerated
+  labels/views/placement are unchanged.
+- **Authored chamfer tolerances now render** (#724 / #698): `render_chamfers`
+  consumes the planner's `DimensionGroup` (chamfer joins `_CONVENTION` as a
+  leader), so a `decorations`-authored ± / limit tolerance on a chamfer leg
+  reaches the placed callout (`C12 ±0.2`). Previously the pass formatted raw
+  feature fields and silently dropped it — the latent #629 bug class the ADR
+  0015 bypass list documents. Untolerated chamfer labels/views are unchanged.
+- **`model.chamfer(face=…)` rejects a non-planar face** (#704): the declared
+  front-end silently accepted a curved face (`normal_at()` does not fail on
+  one) and read garbage legs off it — e.g. a fillet or countersink face became
+  a bogus `ChamferFeature`. It now raises the documented `ValueError`, using
+  the recogniser's own planarity gate (`classify_bevel`), so declared and
+  detected chamfers classify identically by construction.
+
+### Fixed
+
+- **CTC-02/04 no longer drop step-height / overall-height dims** (#733): when
+  the height ladder joined the corridor solve (#689) it silently moved from
+  "places early" to "registers early, places at the drain" — so the immediate
+  machined-feature leader callouts (pockets on CTC-04) filled the front-right
+  strip first and the FORCED principal dims hard-dropped
+  (`placement_unsatisfiable`), a priority inversion the old stage ordering had
+  prevented implicitly. The leader-callout stages (chamfers/fillets/flats/
+  pockets) now sit **after the corridor drain** in `_PASS_SEQUENCE` (joining
+  grooves, which already placed post-drain for exactly this reason): principal
+  dims win by construction, and a callout with no clear room yields with a
+  warning, never a principal-dim error. Callout leaders may shift (they now
+  route around the drained dims); on the `hex_bar` golden this *recovers* a
+  silently starved `m_env_width` envelope dim.
+
+### Removed
+
+- **Three orphaned root exports** (#704): `draftwright.recognise_face_levels`
+  (one recogniser of ~12 — incoherent as a top-level surface; import it from
+  `draftwright.recognition`), `draftwright.dedup_diams` (an analysis internal),
+  and `draftwright.fix_svg_page_size` (still available from `draftwright.export`
+  and the `draftwright.make_drawing` facade). None had any internal or test
+  consumer via the root API.
+
+### Deprecated
+
+- **The seven `Drawing` compat aliases** (`_named`/`_anno_view`/`_pinned`/
+  `_build_issues`/`_pattern_callouts`/`_patterned_holes`/
+  `_dropped_callout_diams`) are formally on notice (#699 slice c, tracked by
+  #720): removal target **0.4.0**, per ADR 0005 §4. Read through the public
+  surface instead — `annotations()`, `iter_annotations()`, `get_annotation()`,
+  `view_of()`, `registry.pinned_names()`/`is_pinned()`, and the new
+  `registry.issues`. All external production and test call sites in this repo
+  are already redirected (`drawing.py` internals ride the aliases until the
+  #720 deletion).
+- **`draftwright.sheet_dsl`** (belated announcement — the alias shim shipped
+  in 0.3.1 when #640 renamed the module to `draftwright.sheet`): importing it
+  warns; it will be removed in **0.4.0**. Import from `draftwright.sheet`, or
+  just `from draftwright import Sheet`.
+
+## v0.3.3 — 2026-07-18
+
+The **layout-quality release**: the occupancy model gets honest about ink shape,
+the last solver-invisible pass joins the corridor solve, and crossed labels shift
+themselves clear. Also completes the #635 consolidation epic (ADR 0005 fully
+executed; ADR 0009's no-invisible-occupant guarantee now holds by construction
+for every auto-pass).
+
+### Added
+
+- **Witness-crossing label shifts** (#690): a dimension label crossed by another
+  dimension's witness stroke now shifts along its own line to clear it — a
+  deterministic post-drain pass using the repair machinery (destination-aware,
+  axis-aligned only, pinned dims exempt).
+
+### Changed
+
+- **L-shaped occupancy** (#685): placement obstacles decompose into per-stroke
+  boxes + label box (helpers ≥0.14 `segments`) instead of one hull, with
+  preset-aware arrowhead pads. View-corner contention between perpendicular
+  strips disappears by construction (a plate-thickness dim returns to its
+  natural strip); packing tightens. The cleanliness ratchet mirrors the model —
+  only transverse stroke crossings are exempt, so parallel overprints and label
+  contacts are detected honestly (twelve hull-artifact allowlist entries burned
+  down; the legitimate shared-datum running-dimension pairs are documented).
+- **Height ladder joins the corridor solve** (#636): step heights + overall
+  height are corridor candidates (the leapfrog witness cursor survives as a
+  deterministic build-time chain; the overall height stacks outermost by
+  construction and packs tighter). An explicitly-requested crowded-step detail
+  view now *deliberately* demotes the overall-height dim when the sheet cannot
+  hold both (previously the carve dropped it silently); a Y-drilled hole's Z
+  location joins the front running ladder per the documented routing.
+- **One typed `BuildState`** (#639): the drawing's build context (analysis,
+  part model, lint geometry caches) consolidates into a single object filled at
+  one builder site, single-writer-guarded; the render passes make zero private
+  `Drawing` reads (fail-closed empty allowlist).
+
+### Fixed
+
+- A prismatic detail view no longer loses both the detail *and* the height dim
+  when space is tight (transactional demotion).
+- The GD&T alternate-strip fallback defers until every corridor has drained, so
+  it can never preempt a sibling strip's reserved corner.
+
+## v0.3.2 — 2026-07-18
+
+The **performance release**: the #602 perf loop (six PRs) plus the
+build123d-drafting-helpers 0.14 upgrade. A dimension-dense drawing now builds
+roughly **14× faster** than v0.3.1 (grid-plate benchmark 40 s → 2.8 s; NIST
+CTC-02 ~107 s → well under half); the full test tier dropped ~14 min → ~3.5 min.
+
+### Performance
+
+- **Placement measures, then builds** (#678, #679): corridor/strip candidates are
+  evaluated on analytical footprints (`dim_footprint`) or a single probe box; an
+  accepted candidate builds its OCC geometry exactly once and re-validates the
+  real box. Dimension constructions on the scattered-plate benchmark: 59 → 21
+  (= the placed count).
+- **Detect once** (#682): the sizing model `_analyse` builds pre-scale is stored
+  on `Analysis.model` and reused as the render model — restoring ADR 0008
+  Amendment 5 ("one inventory, detected once"). ~13 s off CTC-02.
+- **Fillet recognition de-quadratised** (#683): edge→faces adjacency map replaces
+  the O(faces²×edges²) `IsSame` sweep (3.7 M calls on CTC-02); 10.8 s → 4.6 s,
+  byte-identical records.
+- **Lint bounding boxes memoised** (#681): `Drawing.lint()` 0.88 s → 0.07 s per
+  call (identity+location-token-checked cache beside the #143 view-edge cache).
+- **DXF viewport from the known page window** (#680): skips `ezdxf.zoom.extents`'
+  per-entity walk; ~2.4 s per DXF export.
+- **helpers 0.14** (#684): boolean-free dimension ink (upstream #177, filed and
+  fixed same-day) — ~25 ms per `Dimension`, was ~234 ms.
+
+### Changed
+
+- **Tight-span dimensions now render honestly** (helpers 0.14): ISO outside
+  arrows where 0.13 silently degraded to a bare line. Ink extents grow on such
+  dims; placement models the flip analytically. Two placement-golden fixtures
+  re-blessed (ink extent only).
+- Diameter-row leader labels space by **measured text width** and repair
+  helpers' shelf-flip collisions (crowded rows only; ordinary rows unchanged).
+- A plate-thickness dim whose home strip corner is structurally contested falls
+  through to the opposite strip / side view instead of dropping (#684; the
+  underlying one-box-occupancy artifact is tracked as #685).
+
+### Fixed
+
+- `recognise_fillets` neighbour tie-break is order-independent (#683 review).
+- Lint's annotation-box cache prunes departed objects and detects in-place
+  relocation (#681 review).
+
+## v0.3.1 — 2026-07-17
+
+Also lands the **#635 consolidation epic** (P1–P6, all six children closed): the render
+layer's state bus is retired (`annotations/` reads zero `Drawing` privates,
+machine-enforced), the three complexity hotspots are split (nesting ≤2), the whole-package
+import DAG + a test-import ratchet are enforced, and the test suite gains a Hypothesis fuzz
+tier + relational cross-kernel placement gates. All internal / behaviour-preserving
+(byte-identical golden) — no output change.
+
+### Added
+
+- **PNG raster export** — `--format png` on the CLI, or `Drawing.export(formats=("png",))`.
+  Rendered SVG → PDF → PNG via **pypdfium2** (Apache-2.0, Google PDFium BSD-3) + **Pillow**
+  (HPND) — both permissively licensed and pure-wheel, so PNG works cross-platform with **no
+  native cairo** (ADR 0006) and keeps the render path dual-license-friendly. `dpi=` sets the
+  raster resolution.
+
+### Changed
+
+- **Unified export API** — `Drawing.export(out, *, formats=("pdf",)) → {format: path}` is now
+  the single entry point over `svg`/`dxf`/`pdf`/`png`; it returns a dict and internally handles
+  the SVG→PDF→PNG intermediate chain (writing/cleaning up intermediates it wasn't asked to keep).
+  The CLI `--format` gains `png`.
+
+### Deprecated
+
+- `Drawing.export_pdf()` — use `export(formats=("pdf",))["pdf"]`; the method now warns.
+- The legacy `export(svg=, dxf=)` boolean keywords and their `(svg_path, dxf_path)` tuple return
+  — kept for back-compat; prefer `formats=[...]` (the dict API).
+
+## v0.3.0 — 2026-07-14
+
+**Breaking — feature-recogniser API renamed** (ADR 0013 / #568): *deliberate* breaks with
+**no compatibility aliases**, hence the minor bump. Only import paths and callable
+signatures change — but note this release also **changes drawing output**: new machined-
+feature callouts, corrected leader anchors, and a reoriented isometric (see Added / Fixed).
+
+### Added
+
+- **Machined-feature recognition — the #148 epic.** #135 deliberately handled only enclosed
+  rectangular through-slots; this broadens recognition, and each feature is recognised **and**
+  dimensioned **and** authorable (recognise + emit + declare, ADR 0011):
+  - **Blind slots + pockets** — floored recesses, `W × L × D DEEP` (#603).
+  - **Cross / intersecting slots** collapse into single continuous channels (#604).
+  - **Machined flats** on round stock — `{across} A/F` (#605).
+  - **Turned / circlip grooves** — `{width} WIDE × ø{floor}` (#606).
+  - **Slots / pockets in curved stock** — arc-bounded walls (#607).
+  - **Radiused-end (obround) slots** now report their **overall** length (#613).
+- **Chamfer callouts** — `C{leg}` / `{leg}×{angle}°` off each bevel (#560).
+- **Fillet radius callouts** — `R{radius}` off each rounded edge (#561).
+- **Plate / wall thickness** dimensions on multi-plate prismatics (#559).
+- **Prismatic step-position** dimensions locating each shoulder (#555).
+- **Countersink callouts** — `ø6 THRU ⌵ ø14 × 90°` (major-ø + included angle), like the
+  counterbore `⊔ ø.. ↧..` (#558). The countersink rides on `HoleRecord`/`HoleFeature`, so
+  grouping and the callout-width layout estimate account for it.
+- **Every recognised feature round-trips all three ADR-0011 surfaces (#574).** Beyond
+  detection, each feature now **emits** into the generated `--script` Sheet and is
+  **declarable** on the fluent `Sheet` surface — authored from scratch or read off a
+  build123d object: chamfer (#576), countersink (#575), plate (#577), step-position (#585),
+  fillet (#561), plus all the #148 features above.
+- **`draftwright.score.feature_census(part)`** — a per-kind recognised-feature census, a
+  measurement tool for tracking recognition coverage (#608).
+
+### Changed
+
+- **Feature recognisers renamed `find_`/`analyse_` → `recognise_`** and their
+  tuning/dependencies made keyword-only: `find_holes`→`recognise_holes`,
+  `find_bosses`→`recognise_bosses`, `find_hole_patterns`→`recognise_hole_patterns`,
+  `find_slots`→`recognise_slots`, `find_plates`→`recognise_plates`,
+  `find_chamfers`→`recognise_chamfers`, `find_turned_steps`→`recognise_turned_steps`,
+  `find_step_shoulders`→`recognise_step_shoulders`, `analyse_face_levels`→
+  `recognise_face_levels` (the last was also re-exported at top level as
+  `draftwright.analyse_face_levels` → now `draftwright.recognise_face_levels`).
+  Cylinder-analysis substrate (`analyse_cylinders`, `full_cylinders`,
+  `feature_diameters`) keeps its names.
+- **`recognise_step_shoulders` now returns `list[StepShoulder]`** (a frozen dataclass)
+  instead of raw `(axis, position)` tuples; `levels` is keyword-only.
+- **`recognise_turned_steps` now returns `list[TurnedStep]`** (empty for a non-turned
+  part) instead of `TurnedProfile | None`. `TurnedStep` gained an `axis` field (each step
+  is now a self-contained record); `TurnedProfile` remains only as a pipeline aggregate
+  (`TurnedProfile.from_steps(steps)`), no longer a recogniser return.
+- **`recognise_face_levels` now returns `list[FaceLevel]`** (a frozen `FaceLevel(z)`
+  record) instead of `list[float]`.
+- **Recognition record classes renamed to avoid clashing with the IR `Feature` types:**
+  `draftwright.recognition.HoleFeature` → `HoleRecord`, `BossFeature` → `BossRecord`
+  (the IR `draftwright.model.ir.HoleFeature`/`BossFeature` are unchanged).
+- **Page/scale sizing now reads the IR, not recognition records (ADR 0008 / #588).** The
+  part model is built up front and drives sheet sizing, so detected and declared parts share
+  one path and no recogniser record reaches the sheet estimators. Byte-identical to the old
+  estimators except a hole pattern that shares a machining spec with loose holes now sizes
+  separately (no phantom merged `N×`), which can shift a tightly-packed such part's layout.
+
+### Fixed
+
+- **Isometric view orientation (#620).** The iso camera viewed the *rear* (+Y) side against a
+  front-view orthographic set, mirroring asymmetric features; it now views from the same
+  front (−Y) / plan (+Z) / right (+X) combination.
+- **Chamfer leaders (#621)** anchor on the bevel-face centroid (the diagonal middle), not the
+  supporting plane's parametric origin (which projected to a corner/endpoint).
+- **Fillet leaders (#622)** anchor on the radius surface, not the face bounding-box centre
+  (which sat off the round, near the virtual sharp corner).
+- **Authored `ref_pts`-only dimensions** now render (with Z-strip sizing) (#562).
+- **Pitch-dim labels** are cleared off centrelines / bolt circles at creation time, not via
+  the repair loop (#129).
+
+### Removed
+
+- **`draftwright.features` compatibility module** — deleted. Import `recognise_slots`/
+  `Slot` from `draftwright.recognition` instead.
+
+## v0.2.13 — 2026-07-09
+
+**Compose-then-pack is now the layout authority.** This patch finishes the
+ViewBlock footprint migration: estimated and measured layout paths now use the
+same composed block model, furniture declarations reserve their real footprint,
+and measure/repack iterates until stable instead of trusting one measured pass.
+
+### Changed
+
+- **ViewBlock footprints now drive the orthographic stack.** The old
+  estimator-only plan-view balloon lift is gone; plan balloon headroom is part
+  of the plan block in both estimated and measured/repack layout. (#112)
+- **Measure/repack iterates to a fixed point.** Repack now keeps measuring and
+  repacking while view-owned footprints continue to grow, with bounded
+  convergence and tests for multi-pass growth. (#302)
+- **Estimated view-block composition is explicit.** Scale/page fitness, repack,
+  and furniture placement share the same composed block semantics instead of
+  rebuilding corridor estimates independently.
+
+### Fixed
+
+- **Declared callout furniture reserves layout footprint.** User-declared
+  furniture is now measured and fed into the layout model, reducing post-hoc
+  collisions from declarations that previously existed only at render time.
+- **Annotation-box composition owns strip sizing.** Strip depths reduce from
+  composed annotation boxes, making footprint sizing less dependent on divergent
+  scalar estimates.
+
+### Documentation
+
+- **Layout architecture docs reflect the current state.** ADR 0004 and the
+  layout primer now mark the core layout-authority tranche complete and separate
+  remaining work into hardening, coverage, detail-view, and manual-intent phases.
+
+## v0.2.12 — 2026-07-09
+
+**Sheet scripts now round-trip authored layout intent.** This patch finishes the next
+slice of the layout-authority work: below/right ladders, pinned edits, height ladders,
+detected envelopes, and AP242 authored dimensions now survive the declarative Sheet path
+without being invalidated by post-hoc placement or raw imported metadata.
+
+### Added
+
+- **Pinned edit intents route through the corridor solve.** `locate(..., pin=True)` and
+  pinned dimensions now become priority-ranked candidates in the shared corridor instead
+  of fixed post-hoc edits, so user intent participates in the same ordering/spacing model
+  as automatic dimensions. (#511)
+- **AP242 dimensional PMI lowers to authored drafting dimensions.** Imported AP242
+  size/location dimensions now become `AuthoredDimension` IR and generated Sheet scripts
+  emit `sheet.dimension(...)`; unsupported PMI records remain explicit raw fallbacks.
+  (#503, #422, #62)
+
+### Changed
+
+- **`place_dim` is deprecated.** Manual dimensions should use pinned candidate/dimension
+  intent instead of the old incremental edit path.
+- **Prismatic height ladders and detected envelopes round-trip through Sheet scripts.**
+  Generated scripts now preserve `StepLevelFeature` and detected `EnvelopeFeature`
+  fallbacks, preventing CTC01-style ladder swaps and raw STEP-envelope remeasurement
+  drift.
+
+### Fixed
+
+- **Below/right corridors now share one solve.** Side-hole locations, height ladders,
+  PMI/GD&T, envelope dims, and related below/right strip occupants negotiate in the shared
+  corridor rather than competing through separate late passes. (#477)
+- **Sheet script parity is tighter.** Generated scripts preserve member positions, step
+  levels, AP242 authored dimensions, raw PMI fallbacks, and detected envelopes so
+  direct-vs-script CTC01 output has matching annotation names.
+
+## v0.2.11 — 2026-07-08
+
+**Automatic layout now has one authority.** Page/scale selection, section placement,
+furniture footprints, and table/balloon escalation now negotiate through the same layout
+fitness model instead of using separate fixed offsets or first-fit policies after the main
+solve.
+
+### Fixed
+
+- **Solver placement paths are more consistent.** STEP PMI dimensions now queue into
+  the shared ADR 0009 corridor solve instead of carving after the drain; front-view
+  hole callouts use the strip solver instead of fixed row stepping; pitch-dimension
+  fallback searches bounded, obstacle-aware positions; repair no longer hides
+  `annotation_overlap` with a fixed-step nudge; and step-count sizing now handles
+  non-convergence conservatively. (#524)
+- **Page/scale fitness is shared across initial selection and repack.** Later layout
+  passes now compare candidates with the same model as the page chooser, preventing a
+  nominally "better" repack from invalidating the original layout decision. (#519)
+- **Section A-A participates in layout selection.** Section placement is measured as
+  part of page/scale fitness, so a section view is no longer a fixed-offset afterthought
+  that can disappear on dense sheets. (#515)
+- **Furniture placement reserves full rendered footprints.** Section arrows, detail
+  views, balloons, hole/data tables, and other furniture now reserve/check their true
+  rendered footprint instead of just their labels. (#518)
+- **Hole/data tables and balloon rings escalate through the layout model.** Dense table
+  and balloon outputs can move or negotiate for available room rather than dropping from
+  greedy first-fit placement when a small adjustment would fit. (#516, #517)
+
+## v0.2.10 — 2026-07-07
+
+**Declarative-surface fidelity and layout-engine unification.** Editing and re-running a
+declarative `Sheet` script now reproduces more of the drawing (title block, layout, PMI on the
+declared path), a new lint catches stale declarations, and an intentional scale is respected.
+Internally the annotation placer moves further onto the single collect-then-solve (ADR 0009),
+and `kiwisolver` is dropped as a dependency.
+
+### Added
+
+- **Declaration-vs-geometry reconciliation lint** — `lint()` now flags a declared cylindrical
+  feature (hole/boss/step) with **no matching geometry** in the part (`declared_feature_absent`):
+  a callout drawn over empty space because the part was edited but the declaration went stale.
+  Closes the last gap in the "did my edit break the drawing?" loop (over-declaration; coverage
+  lint already caught under-declaration). Matches on axis + ⌀ + in-plane position + bore/boss
+  polarity. Gated on a caller-supplied `model=` (detection can't over-declare). (#487)
+- **Title-block + layout aspects on the `Sheet` DSL** — `Sheet(drawn_by=…, tolerance=…, scale=…,
+  page=…)`; the generated `--script` reproduces them, and the CLI forwards
+  `--drawn-by`/`--tolerance`/`--scale`/`--page` on the sheet path (no more inert-flag warning). (#474)
+- **PMI reproduced on the declared-model path** — `build_drawing(path, model=…, pmi="annotate")`
+  synthesises the STEP AP242 PMI into the declared model, so a declared build draws the same PMI
+  dimensions as the detection path. (#472)
+
+### Changed
+
+- **An intentional explicit `scale=` below the legibility floor is honoured with a warning**
+  rather than rejected. A part deliberately drawn at 1:1 (or `Sheet(scale="1:10")`) whose smallest
+  feature falls under the 10 mm legibility floor now renders (annotations may crowd) instead of
+  raising `ValueError`; only a genuinely degenerate scale (< 0.1 mm projected, where OCCT arcs
+  collapse) is rejected. The floor now binds only the *automatic* scale. (#489)
+- **`kiwisolver` is no longer a dependency.** The 1-D strip solve delegates to the built-in
+  deterministic minimum-total-leader-length PAVA algorithm, which supersedes the Cassowary
+  solver — same placement contract, one fewer third-party dependency. (#507)
+
+### Fixed
+
+- **Dimension placement is more robust under contention.** Candidate **priority** is plumbed
+  through the shared corridor solve, so an authored GD&T frame is kept over a lower-value
+  automatic dimension when a strip is over capacity (#357). Rotational concentric-bore leaders
+  are now bounded to the front-view height with ranked drops, and pitch dimensions are placed
+  onto the obstacle-aware per-side zone-strip solve — retiring the last fixed-offset,
+  count-based stacking placers (the shape behind earlier dense-sheet overruns). (#374)
+
+## v0.2.9 — 2026-07-06
+
+**Declarative GD&T, datums, and surface finish (ADR 0011 Phase 2b/2c).** The `Sheet` DSL can
+now author the drawing information geometry can't carry — geometric tolerances, datum feature
+symbols, and surface-finish marks — by pointing at a build123d feature or face. They render
+through the auto-layout engine as first-class ADR 0009 corridor candidates, placed and spaced
+crossing-free alongside the automatic dimensions (not a post-hoc overlay).
+
+```python
+sheet.datum("A", base_face)
+sheet.control(bore).position(0.1, to="A", diameter=True).perpendicularity(0.05, to="A")
+sheet.diameter(journal).finish("0.8")
+```
+
+### Added
+
+- **Feature control frames (ISO 1101)** — `sheet.control(ref)` returns a chainable builder with
+  one method per **all 14** characteristics (`position`, `flatness`, `perpendicularity`,
+  `cylindricity`, `circular_runout`, `total_runout`, …); `to=` names the referenced datum(s),
+  `diameter=` prefixes the ⌀ tolerance zone, `modifier=` adds a material-condition symbol (Ⓜ/Ⓛ/Ⓟ).
+- **Datum feature symbols (ISO 5459)** — `sheet.datum("A", face_or_feature)`.
+- **Surface finish (ISO 1302, Ra)** — `.finish("1.6")` on a hole / diameter / step handle, or
+  `sheet.finish(ra, face)`.
+- The target view + strip side are **derived from the referenced geometry** (a feature's axis →
+  face-on view, a planar face's normal → edge-on view); `view=`/`side=` override.
+- Render core: `ControlFrame` / `DatumRef` / `Finish` IR features placed by `render_gdt` as
+  first-class corridor candidates, with the real-footprint plumbing needed to space wide frames.
+
+### Fixed
+
+- A control frame referencing an **undeclared datum letter** now warns at build.
+- GD&T placement is **title-block aware** and **falls through to the opposite strip side** before
+  dropping a frame, so stacked control frames place robustly rather than vanishing or overlapping
+  the title block.
+- An invalid glyph spec / degenerate leader in a caller-supplied model **drops the one item with a
+  warning** instead of crashing the whole build (the IR is public input).
+- Turned-shaft rotational furniture now reproduces correctly through the declared-model path
+  (`model=` / `Sheet`) (#476).
+
+## v0.2.8 — 2026-07-06
+
+**`--script` now defaults to the declarative `Sheet` DSL.** The editable script the CLI
+writes is the beautiful-Python surface — one commentable line per feature — for both STEP and
+build123d-object input.
+
+### Changed
+
+- **`draftwright … --script` now emits the declarative `Sheet` script by default** (was the
+  imperative edit-verb reconstruction). The imperative reconstruction is still available via
+  `--script --style imperative`. For prismatic parts the generated script reproduces the same
+  drawing as a direct build; a turned shaft's centre lines and base-diameter style are not yet
+  identical (the remaining validity gap, tracked separately). `--style imperative` now errors
+  clearly on a `module:attr` object spec (it reads a STEP file; use the default `sheet`).
+  The Sheet DSL doesn't yet model the title-block / layout aspects the imperative script
+  embeds, so `--script --style sheet` **warns** when `--drawn-by`, `--tolerance`, `--scale`
+  or `--page` is set rather than silently ignoring it (use `--style imperative` to embed them).
+
+## v0.2.7 — 2026-07-05
+
+**The three authoring modes (ADR 0011 Amendment 1).** Naming how you drive draftwright —
+*just do it* / *auto then tweak* / *generate an editable script* — and filling in the third:
+a declarative `Sheet`-DSL generator, and number-free authoring against the build123d objects
+you built.
+
+### Added
+
+- **`draftwright part.step --script --style sheet`** — generate an **editable declarative
+  `Sheet` script** from a detected part: one commentable line per feature (hole / diameter /
+  step / slot / pattern / envelope), views noted, the auto-section flagged. Comment a line out
+  to drop that feature; the script re-runs to the drawing. The existing `--script` (imperative
+  edit-verb reconstruction) stays the default (#461).
+- **Object-reading `.cbore(tool)` / `.spotface(tool)`** on a declared hole — read the
+  counterbore/spotface's ⌀ **and** depth off the tool object + part (depth measured from the
+  local open face, so a rib/wall elsewhere doesn't skew it). A `Sheet`-declared drawing can now
+  carry **zero magic numbers** and track the geometry parametrically (#462).
+- **`Sheet.of(feature)`** — a fluent handle onto an **existing / detection-seeded** feature (by
+  the build123d object, an index, or the `Feature` itself), so you can `.fit(...)` /
+  `.tolerance(...)` a feature from `Sheet.from_part()` without re-declaring it (#463).
+
+### Docs
+
+- **ADR 0011 Amendment 1** — the three authoring modes and the mode-3 generation surface;
+  records the decision to emit **honest detected numbers** for a detected part rather than
+  fabricate a build123d reconstruction that misrepresents the geometry (#464).
+
+## v0.2.6 — 2026-07-05
+
+**ADR 0011 — the IR as a public input: declare features, don't only detect them.**
+Detection stays the default, but the caller can now *supply* the feature model, so a
+part you built parametrically reads as its own drawing and misdetection becomes
+recoverable by construction. Plus the first Phase-2 aspect layer (tolerances + fits).
+
+### Added
+
+- **`build_drawing(part, model=…)`** accepts a caller-supplied `PartModel` (or a
+  `Sequence[Feature]`); when given, **detection is skipped** and the auto-pass dimensions
+  the declared features. Detection and declaration are two producers of the same IR (#447,
+  ADR 0011 Phase 0).
+- **Object → feature constructors** `draftwright.model.hole` / `boss` / `step` / `slot` /
+  `pattern` / `envelope` — read a feature's geometry off the build123d object you built (⌀
+  from the cylindrical face, axis/location from the bbox), or supply explicit values (#447).
+- **The fluent `Sheet` façade** (`draftwright.Sheet`) — reference the objects you built,
+  declare their drawing aspects, `.build()` / `.export()`. `Sheet.from_part()` seeds the
+  hybrid override mode from detection (#447, Phase 1).
+- **Toleranced dimensions** — a `±` / limit tolerance on a diameter, step, or hole bore,
+  rendered on both the linear and ⌀-callout paths (`Sheet.tolerance(...)`, or a
+  `decorations=` side-layer) (#28, Phase 2 P2a).
+- **Fit-class deviation** — `Sheet.fit("H7")` resolves an ISO 286 fit class to its limit
+  deviation for the feature's nominal ⌀, rendered as the class code (`ø20 H7`, default) or
+  the signed deviations (`show="deviation"` → `ø20 +0.021/0`). Common classes; fails loud
+  outside its table (#29, Phase 2 P2a.2).
+
+### Changed
+
+- **Object constructors honour explicit overrides** — a passed object supplies *defaults*;
+  each explicit keyword overrides that field independently, and invalid public input fails
+  at declaration with a clear `ValueError` (#451, #452).
+- **A declared hole/pattern renders at its declared position** even where detection missed
+  it — the callout membership is sourced from the declared model, not only detection (#448).
+- **`Sheet.model()` / `Sheet.from_part()`** no longer build a full drawing just to return
+  the IR — feature inspection/seeding is now a cheap, no-render path (#453).
+
+### Fixed
+
+- A narrow diameter band hidden under a larger OD silhouette is no longer silently
+  undimensioned (the two feature inventories agreed) (#298).
+
+## v0.2.5 — 2026-07-04
+
+**The editable write API and record-then-finalize.** A detected drawing became an editable
+object, and a generated `--script` became a runnable reconstruction that reaches auto-pass
+quality.
+
+### Added
+
+- **`dwg.model()`** exposes the detected `PartModel` as a read surface, plus feature-
+  referenced add verbs `dimension()` / `callout()` / `locate()` / `furniture()` /
+  `section()` and `drop(feature)`, with a machine-checked completeness audit (#400).
+- **Record-then-finalize** (#426): the verbs record intents in deferred mode and
+  `dwg.finalize()` (auto-run by export) drains them through the auto-pass's own solvers, so
+  a reconstruction reaches auto-pass quality; `--script` now emits a runnable detect-only
+  reconstruction.
+
+### Changed
+
+- **Dimension-line spacing now follows ISO 129-1 / ASME Y14.5 convention** (#347): a wider
+  first-line gap (8 → 10 mm) and tighter, uniform parallel stacking (between-line clear gap
+  4 → 2.5 mm), with the inter-view corridor widened in step. Re-drifts every drawing.
+
+## v0.2.4 — 2026-07-03
+
+A follow-up patch on the ADR 0009 placement rebuild in 0.2.3: it finishes unifying
+the shared "above-view" dimension corridor, adds a layout-overflow safety net, and
+makes two more drop paths non-silent.
+
+### Changed
+
+- **Plan-view X location dimensions, side-view Y location dimensions, and a
+  coincident slot-position dimension now share one collect-then-solve pass** (ADR
+  0009 Amendment 6, #345/#346). Previously each pass carved the strip independently,
+  so a hole location and a slot position measuring the same datum span could both be
+  drawn, and the location ladder could come out non-monotonic. One solve now dedups
+  the coincident span (keeping the higher-priority location dimension) and orders the
+  whole ladder as segregated, monotonic runs — feature-size dimensions nearest the
+  view, datum locations nesting outward by distance.
+
+### Fixed
+
+- **`choose_scale` never returns an overflowing layout** (#350). Scale selection
+  could pick a scale whose composed block layout exceeded the drawable area; it now
+  rejects any overflowing candidate.
+- **A hole location and a coincident slot position are no longer drawn twice** (#345),
+  including at fractional datum distances where a display-value snap gap previously let
+  the duplicate escape deduplication.
+- **The plan-view location ladder is monotonic** (#346) — running dimensions off a
+  shared datum stack outward in ascending order instead of interleaving.
+- **A dropped balloon is non-silent** (#387). A balloon that cannot be placed now
+  reports the drop and clears its `callout_dropped` state precisely, instead of
+  vanishing with no on-sheet signal.
+
+## v0.2.3 — 2026-07-03
+
+A large patch release: the **annotation-placement engine was rebuilt** as a
+collect-then-solve *boundary-labeling* stage (ADR 0009). Placement is now
+deterministic and minimises total leader length, and the recurring class of
+overlaps where a label was drawn on top of an "invisible" occupant — a leader
+shaft, a witness/extension line, the section hatch — is removed by construction.
+Drawing output changes for many turned, cross-drilled, and multi-feature parts.
+
+### Changed
+
+- **Every annotation in a view's margin is now placed by one collect-then-solve
+  pass** (ADR 0009, #317–#323). Dimensions, hole callouts, turned-diameter
+  leaders, and the section hatch share a single occupancy model instead of several
+  independent passes each blind to the others. When a strip is over capacity it now
+  drops the *lowest-priority* annotation (smallest bore first) rather than whichever
+  pass happened to run last. The legacy strip cursor is retired.
+- **Leader placement minimises total leader length, deterministically** (P4,
+  #318). A per-strip solve places each label at the shortest-leader position that
+  keeps the labels in order and clear of keep-out rows (a view centre-line, a
+  dimension's extension line); central/coaxial callouts are anchored to the
+  view-centre row. Output is reproducible across platforms and Python versions.
+- **`scipy` is no longer a dependency** — the leader solve is a small deterministic
+  algorithm (weighted-median isotonic regression), not a linear program.
+- **Output changes** for turned, cross-drilled, and multi-feature parts whose
+  margin annotations are now positioned by the unified solver.
+
+### Fixed
+
+- **A PMI bore-diameter dimension spans the bore radius, not the full diameter**
+  (#360). A `pmi="annotate"` diameter callout drew its witness lines at ±diameter
+  from the centre — twice too wide, missing the bore edges.
+- **A bore coaxial with a rotational part's turning axis is no longer
+  over-dimensioned** (#309). It carried a redundant offset *and* height location
+  dimension even though its centre mark already locates it.
+- **A dropped turned step-length chain is no longer silent** (#362). When a turned
+  head's shoulders are too crowded to dimension, the drop is now reported
+  (`step_dim_dropped`) instead of vanishing with no lint or on-sheet signal.
+- **A diameter callout can no longer overprint a bore callout's leader shaft**
+  (#358). The turned-diameter column now avoids the *full* footprint of existing
+  annotations, not just their text boxes.
+- **The balloon ring hugs its dimensions on a cramped sheet** (#349) — its band
+  depth is clamped to the drawable area.
+- **Dimension detection is robust to `SafeDimension`** (#335/#349) — the corridor
+  and balloon-ring filters test the dimension *type*, not a class-name string, so a
+  future dimension subclass can't slip through.
+
+## v0.2.2 — 2026-06-30
+
+A patch release of turned-part dimension-placement fixes and a CLI start-up
+speed-up. Drawing output changes for the affected turned/cross-drilled parts.
+
+### Fixed
+
+- **A coaxial bore callout on a *stepped* turned shaft is now lifted off the round
+  view's centre axis** (#305). The earlier fix only triggered for a uniform
+  (`is_rotational`) cylinder; a stepped shaft (e.g. the gramel GRM-03 drive screw)
+  has a turned step profile but isn't classified rotational, so its `⌀… ↓…` bore
+  callout was still leadered straight along the centreline, with the centre mark
+  running through the text. The lift now also fires for a turned-profile part.
+- **A side-drilled hole's location dimension now stacks *inside* the overall
+  envelope dimension** (ISO order — overall dim outermost, feature/location dims
+  nearer the view). It was placed *outside* the envelope, which forced the shorter
+  location dim's arrowheads to flip outward and clash (seen on GRM-01 and GRM-02).
+  The mandatory overall dimension is still guaranteed placement.
+
+### Changed
+
+- **CLI shell completion and `--help` are fast again** (#313). The Typer CLI and
+  the heavy CAD engine are now imported lazily, so tab-completion and `--help` no
+  longer pay a ~6 s engine-import cost; a real drawing run is unaffected.
+
+## v0.2.1 — 2026-06-30
+
+A patch release focused on **turned-part dimensioning legibility**: crowded
+step-length chains and fine turned heads are now drawn legibly instead of crammed.
+Drawing output changes for affected turned parts.
+
+### Added
+
+- **Automatic enlarged detail view for a crowded turned head** (#304). A turned
+  part with a fine cluster of steps near one end and a long shaft (e.g. a thumbwheel
+  drive screw) cannot have its head dimensioned legibly in line at any sensible
+  scale. The head is now located as one block on the main view and broken out into
+  an enlarged **DETAIL A — SCALE n:1** — the textbook treatment — firing
+  automatically when a head's shoulders fall below the page legibility floor.
+
+### Changed
+
+- **Crowded turned step-length chains stagger across two tiers** (#293) instead of
+  cramming or being skipped. When the labels would collide on one line, the ISO
+  129-1 staggered convention alternates them between a near and a far tier so every
+  step length stays legible at the drawing's own scale — no rescale needed. A roomy
+  chain stays on a single tier.
+- **Detail views are now one unified pipeline** (#307). The prismatic step-height
+  detail and the new turned-head detail flow through a single
+  detect → request → render path; several crowded regions become DETAIL A/B/…
+- **Output changes** for turned parts whose step chains were previously crammed, or
+  whose fine heads are now broken out into a detail view.
+
+### Fixed
+
+- **A coaxial bore callout no longer overlaps the round view's centreline** (#305):
+  its leader is angled off the centre axis so the callout text sits in clear space.
+
+### Internal
+
+- A new **layout-cleanliness invariant test** asserts that finished drawings have no
+  view/annotation collisions across representative part archetypes, and the
+  measure-and-repack pass gained a trigger for an annotation growing into a
+  neighbouring view's line-work (so the views spread to make room).
+
+## v0.2.0 — 2026-06-30
+
+A major release. draftwright took ownership of feature recognition and linting
+(ADR 0007) and was re-architected onto a feature-IR + dimensioning-planner
+"compiler" (ADR 0008), gaining a Typer CLI and a portable, pure-Python PDF path
+along the way. **Generated drawings change** versus v0.1.13: the new pipeline
+dimensions parts more completely and consistently, so placement and the set of
+dimensions can differ — output is not byte-compatible with prior releases.
+
+### Changed
+
+- **Re-architected onto a feature-IR + dimensioning planner** (ADR 0008). The
+  engine is now a compiler: detectors build one feature inventory → a typed IR /
+  `PartModel` → a dimensioning planner emits render-intents → shared
+  layout/projection/export. Orientation and feature-kind are *data in the IR*,
+  not code branches, and every feature class (holes, patterns, counterbores,
+  slots, turned diameters/steps, centre marks, location dims, envelope/OD,
+  section A–A, PMI/GD&T) was migrated onto this one path; the old parallel
+  recognisers and placement passes were deleted as each was replaced. Net effect
+  for users: more complete, more consistent drawings — but output differs from
+  v0.1.13.
+- **draftwright owns feature recognition and linting** (ADR 0007). The hole/
+  boss/cylinder/pattern recognisers, the slot/turned-step recognisers, and the
+  feature-coverage lint engine are vendored into `recognition/` and `linting.py`;
+  `build123d-drafting-helpers` is now purely the rendering library.
+- **The CLI writes a PDF by default** and takes a `--format` selector (#288).
+  Previously `draftwright part.step` emitted SVG + DXF; it now emits a single PDF.
+  Choose outputs with `--format` (a comma-list, with an `all` alias) —
+  `--format pdf,dxf`, `--format svg`, `--format all`. The library API is
+  unchanged: `make_drawing(...)` / `Drawing.export()` still write SVG + DXF.
+- **PDF export is now pure-Python and a core capability** (#288). The renderer
+  moved from `cairosvg` (which `dlopen`s the native `libcairo` system library —
+  absent on stock macOS/Windows, so PDF-by-default would have crashed there) to
+  `svglib` + `reportlab`, both pip-installable wheels with no system dependency.
+  PDF therefore works out of the box on every platform; output is visually
+  identical to the cairo renderer.
+
+### Added
+
+- **A Typer command-line interface** (#289/#291): shell completion
+  (`--install-completion` / `--show-completion`), rich `--help`, and `--version`
+  (reports the installed distribution version). All existing flags are preserved.
+- **`--format` output selector** (#288) — `pdf` (default), `svg`, `dxf`, or `all`,
+  as a comma-list.
+- **Turned-part dimensioning**: axial step-length recognition and chains
+  (#188/#189/#231), step-diameter callouts, collapse of a uniform step
+  staircase to an "N× length" note (#290), and OD of a horizontal (X/Y) round
+  body dimensioned on the profile view (#292).
+- **Slot recognition and dimensioning** converged onto the IR as `SlotFeature`
+  (#242), and **section A–A** is now triggered by the planner (#271).
+- **A Contributor License Agreement** (#183).
+
+### Removed
+
+- **The `--pdf` flag** (use `--format pdf`, the new default), the **`[pdf]`
+  install extra**, and the **`cairosvg` dependency** (#288).
+- The byte-exact golden-output test harness (#190) — regression coverage rests on
+  the geometry-level and standards suites (ADR 0005 §3 / ADR 0007).
+
+### Fixed
+
+- Locate **every** side-drilled (off-axis) hole, not just the first (#225/#286).
+- Don't mis-detect a prismatic part with incidental cylinders as a turned part
+  (#293/#294); drop phantom zero-diameter turned steps (#279/#284); skip an
+  illegibly-dense step-length chain rather than cram it (#293/#296).
+- Degraded hole-pattern callout consistency (#262/#274).
+- Windows `python -m draftwright.make_drawing` CLI smoke / entrypoint (#181/#182).
+
+## v0.1.13 — 2026-06-27
+
+### Changed
+
+- **Requires `build123d-drafting-helpers>=0.13.0`; text pinned to bundled fonts**
+  (#149, ADR 0006). draftwright now vendors IBM Plex (OFL-1.1) and renders and
+  measures all text via `font_path` — IBM Plex Mono for dimensions/callouts/notes,
+  IBM Plex Sans Condensed for the title block — instead of resolving the system
+  font name `"Arial"`. Resolving a name substitutes a different font on Linux,
+  which shifted the whole sheet ~1 mm; pinning a bundled font file makes generated
+  layout **deterministic across Linux/macOS/Windows** and gives a consistent
+  typeface. **Drawing output changes**: positions shift slightly from prior
+  releases and labels render in IBM Plex (helpers #172).
+
+### Internal
+
+- **Compiler-pipeline module split** (#138, ADR 0005). The two large modules
+  `make_drawing.py` (3,907 lines) and `annotate.py` (2,587) were decomposed into a
+  DAG of focused stage modules — `projection`, `sheet`, `analysis`, `drawing`,
+  `builder`, the `annotations/` subpackage (sections/turned/pmi/holes/orchestrator),
+  alongside the existing `registry`/`linting`/`repair`/`export`/`fonts`. Annotation
+  identity, the lint coverage signal, and the deterministic repair loop each gained a
+  single owner; `make_drawing.py` / `annotate.py` are now thin compat facades, so all
+  existing imports and the `draftwright` CLI entry point keep working. A golden-output
+  regression gate verified every step is behaviour-preserving (output byte-identical),
+  and mypy was tightened on the settled contracts. No public API or drawing-output
+  change. (Phases #160–#166.)
+
+## v0.1.12 — 2026-06-21
+
+### Changed
+
+- **Requires `build123d-drafting-helpers>=0.12.0`** (#92, #122). draftwright now
+  consumes the new sub-clustered hole-pattern recognition, the
+  `feature_diameters()` coverage inventory, the persistent `view_edge_cache`,
+  and the `ViewCoordinates.from_viewport()` ISO projection basis.
+- **Grouped hole-pattern callouts** (#92, #111, #114). A recognised perimeter,
+  grid, or bolt circle collapses to a single `n× ⌀ …` callout plus its pattern
+  dimensions instead of a balloon on every hole. A spec group now sub-clusters
+  into multiple patterns (a perimeter → its edge `LinearArray` rows, a filled
+  lattice → one `RectGrid` with a `(rows×cols)` callout and both pitch
+  dimensions); only genuinely unpatterned holes fall back to the per-hole table.
+  On NIST CTC-02 the table shrinks from 61 rows to the unpatterned remainder.
+- **Layout overhaul — compose-then-pack** (#121, #112, ADR 0004). Each view owns
+  the annotations created against it, and the resulting view blocks are packed
+  disjoint with automatic page/scale escalation. This eliminates cross-view
+  overlap — most visibly, plan-view balloons landing on front-view dimensions.
+- **Drawing attribution** (#120). The title block records the author, the SVG
+  and PDF carry a clickable draftwright hyperlink, and a "generated by
+  draftwright" note is written to the SVG/DXF/PDF file metadata.
+- **Gap between wrapped hole-table column blocks** (#123) so a chart that wraps
+  into several blocks reads as distinct columns.
+
+### Fixed
+
+- **Plan-view top balloon ring no longer floats over a phantom corridor** (#125).
+  The hole-table escalation deletes the X-location dimensions but left their
+  stale depth in the strip cursor, so the top balloons were parked far above the
+  view. The ring is now sized to the real dimension stack, so the top-side
+  leaders are short like the other three sides.
+- **No more phantom `feature_not_dimensioned` warnings** on slot-ends and shallow
+  recesses, via the helpers 0.12.0 `feature_diameters()` coverage inventory
+  (#92).
+
+### Internal
+
+- `AnnoBox` box-model footprint foundation and the four-side balloon ring placed
+  in a reserved view halo (#111, #112); the title block is pinned as a
+  first-class layout block (#112).
+
+## v0.1.11 — 2026-06-19
+
+### Changed
+
+- **Feature-coverage lint is assembly-aware.** A general-arrangement drawing of
+  a multi-solid part deliberately omits each part's bores (they belong on detail
+  sheets), so `feature_not_dimensioned` / `feature_count_mismatch` are now
+  emitted at `info` rather than `warning` when the part is multi-solid — out of
+  the warning count and quality score, but still queryable. Auto-detected;
+  override with `build_drawing(..., assembly=True/False)` or
+  `lint_feature_coverage(..., assembly=...)` (#69).
+
+### Fixed
+
+- **`place_dim` now labels the real-world length, not the page distance**, at
+  non-1:1 scale. Previously a dimension placed at a scale other than 1:1 showed
+  the on-page millimetre span instead of the true model dimension (#104).
+
+### Internal
+
+- **`make_drawing.py` decomposed (#98).** The per-view projection math and the
+  analysis namespace were deduplicated and typed (the namespace is now a frozen
+  `Analysis` dataclass), and the annotation passes were extracted into a new
+  `draftwright.annotate` module on top of a shared `draftwright._core`. The
+  module graph is a DAG (`layout → _core → {make_drawing, annotate}`) and
+  `make_drawing.py` shrank from ~5,270 to ~2,930 lines. No public API or
+  behaviour change.
+
+## v0.1.10 — 2026-06-18
+
+### Added
+
+- **Constraint-based layout engine (ADR 0003).** A new `draftwright.layout`
+  module with a `Placeable` protocol and a `LayoutSolver`: a 1D Cassowary strip
+  solver (`solve_strip`, with per-pair gaps) and a 2D free-rectangle placer
+  (`place_box` / `fit_box`) that positions a box in a free part of the page
+  clear of the views, title block, and existing annotations. Hole-callout and
+  turned-diameter placement now run on the solver. The engine grows per real
+  consumer; a monolithic global 2D solve is deferred (see the ADR).
+- **Hole table + balloons (#93).** `dwg.add_table(rows)` places a generic data
+  table (gear data, BOM, revision block, …) in a free corner via `place_box`;
+  `dwg.add_hole_table(view)` builds a hole chart from the detected holes with a
+  circled balloon tag at each hole. A **too-dense plan view now auto-escalates**:
+  a part the layout cannot legibly dimension hole-by-hole is replaced by a
+  complete per-instance hole chart (`TAG | ⌀ | X | Y`, datum-relative) plus
+  balloons, instead of silently dropping callouts and location dims. The chart
+  wraps into multiple column-blocks to fit the page.
+- **External turned diameters (#77).** A turned part lying along the X axis now
+  gets ø leader-callouts for its external stepped diameters, with thread/worm
+  patches collapsed into a single boss.
+- **Pin / manual override (#89).** `dwg.pin(name)` / `dwg.unpin(name)` fix an
+  annotation's position so `repair()` — and the layout engine — never move it; a
+  deliberate (human or AI) placement wins over automatic layout.
+
+### Changed
+
+- Hole-callout and turned-diameter placement is deconflicted through the shared
+  `LayoutSolver` instead of ad-hoc per-pass logic (no output change).
+
+### Fixed
+
+- **Exact circles recovered for revolution silhouettes.** `project_to_viewport`'s
+  HLR returns the on-axis silhouette of a turned feature (or a concentric
+  gear-tooth-tip arc) as an approximating spline, not a true circle — splines in
+  the DXF where CAM expects `CIRCLE`/`ARC`, and fitted rather than exact radii.
+  `add_view` now refits any silhouette whose samples are equidistant from a
+  recognised revolution axis back to an exact circle/arc (#67).
+- **Blind-hole depth no longer measured across solid boundaries.** On a
+  multi-solid assembly, coaxial bores in different bodies were merged into one
+  hole, reporting a depth spanning the inter-body gap (the ⌀9.8 ↓111.4 symptom).
+  Fixed upstream in `build123d-drafting-helpers` 0.10.1; the dependency pin is
+  bumped to `>=0.10.1` to pick it up (#68).
+
+### Docs
+
+- The skill and generated-script header now lead with the domain API
+  (`features` / `place_dim` / `repair` / `lint_summary`) and the
+  build → critique → fix loop. ADR 0003 records the layout architecture; ADRs
+  0001/0002 remain the editing-model and lint→repair foundations.
+
+## v0.1.9 — 2026-06-16
+
+### Added
+
+- **Domain-semantic editing API.** `dwg.features(view)` returns detected holes
+  and features grouped by machining spec in page coordinates, and
+  `dwg.place_dim(p1, p2, side, view, draft, name=…)` places a dimension from
+  domain inputs — the vocabulary a script (or an AI assistant) needs to edit a
+  drawing without hand-computing page geometry (#25, #26).
+- **`dwg.annotations()` and `dwg.get_annotation(name)`.** Introspect what is
+  already on the drawing — a `{name: type}` map and a name lookup — so a script
+  can make incremental edits without risking a silent name-collision replace
+  (#27).
+- **`dwg.view_bounds(view)`.** Returns `(x_min, y_min, x_max, y_max)`, the page
+  bounding box of a view's projected geometry (or `None` for an unknown view),
+  so free-form notes and leader elbows can be placed just outside a view without
+  guessing offsets from `dwg.at()` (#28).
+- **Lint findings carry a suggested fix.** Each repairable lint issue now
+  includes a ready-to-run domain-API call snippet, so acting on a finding is one
+  copy-paste away (#29).
+- **Lint→repair loop.** `Drawing.repair()` — run by default in `build_drawing` —
+  mechanically resolves the lint codes that have a deterministic placement fix:
+  overlapping labels are pushed apart and wrong-side dimensions are flipped. A
+  pass that would net-increase the issue count is rolled back, so repair never
+  makes a drawing worse (#30).
+- **TYP / representative dimensioning for uniform step patterns.** A run of
+  equal-rise, equal-going steps is dimensioned once and labelled representative
+  (TYP) instead of repeating identical dimensions down the ladder (#45).
+- **Enlarged detail view for crowded step clusters (MVP).** When shoulders are
+  too closely spaced to dimension legibly at sheet scale, an opt-in
+  (`detail_view=True`) detail view re-draws them at a larger scale (#42).
+
+### Changed
+
+- **BREAKING: the annotation list `dwg.annotations` is renamed to `dwg.items`.**
+  `dwg.annotations` is now a method (see Added); the ordered, mutable list of
+  annotation objects it used to be is now `dwg.items`. Pre-1.0 with no published
+  consumers, so the clearer name was taken now rather than spelling the new query
+  method awkwardly (#27).
+
+### Documentation
+
+- ADRs 0001 (editing model) and 0002 (iteration loop) record the design
+  direction behind the domain API and the lint→repair loop (#51).
+
+## v0.1.8 — 2026-06-16
+
+### Changed
+
+- **Automatic scale selection now minimises the sheet size.** The preference
+  ladder is page-major: every standard scale on the smallest sheet is tried
+  before the next sheet up, so a part lands on the smallest sheet it fits at the
+  largest scale that sheet allows. A 20 × 15 × 10 mm part is now drawn 2:1 on A4
+  instead of 5:1 on A3 — a smaller sheet is preferred over a larger enlargement
+  scale. Reductions keep their legibility-first balance, so a too-big part is
+  not over-reduced onto a small sheet.
+- **A specified page now enlarges to the best fitting scale.** When the caller
+  fixes the page (`--page A3`) or scale, scale selection packs the isometric
+  view into the largest empty rectangle the placement engine actually uses (it
+  may sit in vertical headroom above the views), instead of charging it a column
+  in the view row. A long, short part — e.g. a 100 × 10 × 11 mm staircase — now
+  fills a requested A3 at 2:1 where it was previously under-scaled to 1:1.
+  Automatic selection (no page/scale given) keeps the conservative row model,
+  which reserves enough room to place every annotation rather than dropping some
+  onto a tighter sheet (staircase review).
+- **Isometric view growth is capped.** The iso is fitted to fill its zone but no
+  longer grows past 1.3× sheet scale; on an oversized sheet it could previously
+  balloon to ~8× and dwarf the dimensioned orthographic views. Shrinking to fit
+  a small zone is unchanged.
+- **Step heights are dimensioned only where legibly separable.** After the
+  adaptive cap (#36), a part with many closely-spaced shoulders (e.g. NIST
+  CTC-02 at 1:5) tried to dimension faces only ~1 mm apart on the page. A step
+  is now dimensioned only if it is both tall enough from the base *and* at least
+  one legible step-height above the previously dimensioned one; the rest surface
+  as `step_dim_dropped` (use a detail view). "Fits" is not the same as
+  "legible" (#41).
+- **Hole-location dimensions are gated for legibility.** A hole-dense part (e.g.
+  NIST CTC-02, ~38 distinct hole locations) previously stacked every location
+  reference into a tall, busy tower above the views — "fits" is not "legible".
+  Each axis's references are now gated by inter-dimension page spacing
+  (`_legible_locations`, analogous to the step-height gate #41): only locations
+  at least one value-label footprint apart on the page are dimensioned; the rest
+  surface as `location_ref_dropped` (full fidelity belongs in a detail view,
+  #42). Sparse parts are unchanged (#43).
+- **Tighter location-dimension tier pitch.** The vertical pitch between stacked
+  X/Y location dimensions is now derived from the label footprint
+  (`font_size + 2·pad_around_text`, ≈7 mm) instead of a looser `font_size·3`,
+  so location stacks pack closer (#41).
+
+### Fixed
+
+- **Phantom step corridor no longer blocks a larger scale.** Page/scale
+  selection reserved a step-ladder corridor sized for *every* candidate
+  horizontal face, including ones the legibility gate would never dimension. A
+  part with many sub-legible faces (e.g. a staircase with 15 tiny treads) was
+  forced onto an oversized sheet at 1:1. Scale selection now iterates so the
+  reserved corridor matches the step count actually placed, freeing the part to
+  pick a tighter sheet (staircase.step review).
+- **Engraved-text faces are no longer dimensioned as steps.** `analyse_face_levels`
+  gained a `min_area_frac` filter; a horizontal face counts as a step only if
+  its area is at least 1% of the part's plan footprint. This drops sub-feature
+  faces (fragments of engraved numbers/text) that were surfacing as phantom
+  shoulders — e.g. a 0.57 mm² digit face dimensioned as z=6.4 on staircase.step.
+- **Overall-height dimension nests outside the step dims.** The overall height
+  is now placed last on the front view's right ladder so it sits outermost, with
+  the step-height dims inside it; extension lines nest instead of leapfrogging
+  (staircase.step review).
+
+## v0.1.7 — 2026-06-15
+
+### Added
+
+- `Drawing.lint_summary()` — a JSON-friendly aggregate of `lint()` for
+  non-interactive callers (scripts, or an LLM via the API): severity counts,
+  per-code counts, a `geometry_issues` tally (standards/geometry checks vs pure
+  layout), a `passed` flag, a coarse 0–1 `score`, and the full issue list. Gives
+  a single signal to gate and optimise on without rendering the SVG (#32).
+
+### Changed
+
+- **Adaptive annotation placement.** The three hard-coded cardinality caps —
+  four hole callouts per view, four hole location references per part, and three
+  step-height dimensions — are removed. The engine now places as many as the
+  available strip/corridor space allows (callouts largest-first, locations
+  nearest-datum-first, every legible step), so a part with room is dimensioned
+  completely instead of dropped to an arbitrary count. An annotation that
+  genuinely doesn't fit is never force-placed; it surfaces via lint
+  (`callout_dropped` / `location_ref_dropped`, warning severity). On the NIST
+  CTC parts this raises coverage substantially (e.g. CTC-02: 4 → 36 location
+  dimensions, 4 → 9 callouts) with no error-severity lint (#36).
+- **No silent annotation drops.** Every place the layout has to drop an
+  annotation now records a machine-readable lint issue, surfaced by `lint()`,
+  so a short drawing always carries a reason. A dropped callout names its
+  diameter and is excluded from `feature_not_dimensioned` (no double-report).
+  `placement_unsatisfiable` (error severity) is reserved for the degenerate
+  case where space was reserved but an annotation still could not be placed
+  (#32).
+- **Layout constants derived from first principles.** Bare, fixture-tuned
+  constants (strip slot widths, callout label widths, isometric fit factor) are
+  now computed from text metrics and page size rather than hard-coded, so the
+  layout generalises to unseen geometry instead of fitting the test cases (#31).
+- `_auto_annotate` clears its build-time lint records on re-entry, and repeated
+  `lint()` calls are stable (#32).
+
+### Fixed
+
+- AP242 / PMI STEP import segfault: STEP geometry is now read directly via
+  `STEPControl_Reader`, avoiding the XCAF/PMI read that crashed (SIGSEGV) on
+  with-PMI files such as NIST CTC-02 (#20).
+
+### Tests
+
+- Overfitting guards pin the general layout behaviour on turned/hybrid parts
+  (flange OD + bolt circle), multi-bore parts, and the step-legibility boundary
+  (#13).
+- The full NIST CTC set (AP203 and AP242) builds and is covered by the slow
+  end-to-end tier.
+
+## v0.1.6 — 2026-06-15
+
+### Fixed
+
+- Section-view boolean cut on cast geometry: the exact `body - Box(...)` boolean
+  raised an uncatchable `Standard_DomainError` (C++ abort, SIGABRT) on some parts
+  (NIST CTC-04), crashing the whole drawing. `_fuzzy_cut()` now runs
+  `BRepAlgoAPI_Cut` with a small fuzzy tolerance and keeps solids-only, making
+  the section cut robust (#20, #22).
+
+### Tests
+
+- NIST CTC-04 (both AP203 and AP242) now build with a clean section view and are
+  covered by the CTC build tests.
+- Known: CTC-02 AP242 still segfaults inside OCCT's AP242/PMI STEP read (#20),
+  excluded from build tests.
+
+## v0.1.5 — 2026-06-15
+
+### Fixed
+
+- CTC-02 spurious full-page line: build123d's `ExportSVG` projected
+  circle-edge-on edges (hole/fillet rims seen edge-on) as elliptical arcs with
+  a near-zero minor radius, which renderers blow up into full-page lines.
+  `sanitize_svg_arcs()` rewrites any arc with a sub-1e-3 mm radius into the
+  straight line it actually is, leaving real arcs untouched (#19). Not a PMI
+  issue — the file is AP203 geometry-only.
+
+### Tests
+
+- Added the full NIST CTC set (01–05) as fixtures, both AP203 geometry-only and
+  AP242 (with-PMI) variants.
+- Heavy end-to-end CTC fixture builds are marked `slow` and deselected from the
+  default `pytest` run (fast normal run, ~4.5 min); CI runs the fast tier across
+  the OS/Python matrix and the slow tier once.
+- Known: AP242 CTC-02 and both CTC-04 variants crash OCCT on import (#20); their
+  fixtures are excluded from build tests.
+
+## v0.1.4 — 2026-06-15
+
+### Changed
+
+- Feature annotations (hole callouts, location dimensions, section view) now
+  fire on feature presence independent of the turned/prismatic classification,
+  so turned-and-drilled parts (e.g. flanges) get both the OD/centreline base
+  set and per-hole callouts plus bolt-circle furniture (#10).
+- Isometric view placement now uses a general largest-empty-rectangle search in
+  place of the wide/flat-on-A3 special case (#11).
+- Concentric bore-leader stacking is generalised beyond three, and the
+  step-height dimension gate is now a single derived constant (#10, #12).
+
+### Internal
+
+- Single-sourced duplicated geometry constants from the draft preset (#12).
+- Minor comment and logging cleanups.
+
+## v0.1.0 — 2026-06-14
+
+Initial release — spun out of `build123d-drafting-helpers` v0.9.1.
+
+The automated drawing engine (`make_drawing`, `build_drawing`, `Drawing`)
+was previously part of `build123d-drafting-helpers`. It is now a separate
+AGPL-licensed package that depends on `build123d-drafting-helpers>=0.9.1`
+for annotation primitives.
+
+### Migration from build123d-drafting-helpers
+
+```python
+# Before
+from build123d_drafting import make_drawing, Drawing, build_drawing
+
+# After
+from draftwright import make_drawing, Drawing, build_drawing
+```
+
+### Features (carried over from build123d-drafting-helpers)
+
+- **`make_drawing`** / **`build_drawing`** — automatic multi-view technical
+  drawing from a build123d solid: view layout, scale selection, orthographic
+  projection, dimension placement, title block.
+- **`Drawing`** — composable drawing object with `.lint()`, `.add()`,
+  `.export_svg()`, `.export_dxf()`.
+- **`choose_scale`** — ISO/ASME standard scale selection.
+- **`lint_feature_coverage`** — checks annotation coverage against detected
+  part features (holes, bosses, bolt circles).
+- **Section A–A views** — automatic section view for blind/stepped holes,
+  with ISO 128-44 solid filled cutting-plane arrows and ISO 128-50 45°
+  hatching on the cut face.
+- **`generate_script`** — generates a standalone drawing script from a STEP
+  file.