testerkit 0.4.0__tar.gz

testerkit-0.4.0/.gitignore

@@ -0,0 +1,39 @@
+.venv/
+__pycache__/
+*.pyc
+*.egg-info/
+dist/
+/data/
+/examples/*/data/
+.ruff_cache/
+.pytest_cache/
+.coverage
+.claude/settings.local.json
+.claude/scheduled_tasks.lock
+.tmp/
+examples/uv.lock
+examples/*/uv.lock
+dist/
+.playwright-mcp/
+report_*.html
+*.mp4
+*.pdf
+
+# Generated skill refs (built from docs/ and models.py)
+testerkit/skills/refs/
+
+# JSON schemas — regenerated via `testerkit schema export` / `testerkit init`
+# / `testerkit schema refresh`. The source of truth is the live Pydantic
+# models in src/testerkit/models/. Don't commit snapshots.
+/schemas/
+.benchmarks/
+
+# Transient logs (daemon, serve, stream, uicheck)
+*.log
+
+# Local databases and env files
+*.duckdb
+*.duckdb.wal
+.env
+.env.local
+.env.*.local

testerkit-0.4.0/CHANGELOG.md

@@ -0,0 +1,559 @@
+# Changelog
+
+All notable changes to TesterKit are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
+this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+Pre-1.0 note: the public API is unstable. Breaking changes are possible in any
+0.x release and will be called out in this changelog.
+
+## [Unreleased]
+
+## [0.4.0] - 2026-07-18
+
+The rebrand release: **Litmus is now TesterKit.** Releases through 0.3.1 were
+published on PyPI as `litmus-test`; from 0.4.0 the package is `testerkit`.
+Version numbering continues unbroken — the rename is cosmetic, the software's
+maturity is not.
+
+### Changed
+
+- **BREAKING** Renamed the entire project `litmus` → `testerkit`. The PyPI
+  distribution is `testerkit` (was `litmus-test`); the import package is
+  `testerkit` (was `litmus`); the CLI is `testerkit` (was `litmus`); environment
+  variables are `TESTERKIT_*` (were `LITMUS_*`); the project config file is
+  `testerkit.yaml` (was `litmus.yaml`); the bundled agent skills are `testerkit-*`;
+  the MCP server is `testerkit`. There are **no aliases** — update imports,
+  scripts, environment variables, and config filenames. The on-disk data-dir key
+  changed to `testerkit`; existing local data under the old key is not migrated
+  (pre-1.0).
+
+### Added
+
+- Class-hoisted instrument fixtures hold their reservation across the whole test
+  class (sequence): a class that `usefixtures` an instrument role reserves it for
+  the lifetime of the class container step — not just per method — via reentrant
+  refcounting, so the lock never drops between two methods of the same class.
+- The operator UI carries the TesterKit brand: wordmark in the sidebar, mark in
+  the header, and a real favicon — replacing the placeholder lightning-bolt emoji.
+  Documentation screenshots were regenerated against the new look.
+
+### Fixed
+
+- Flight daemon thread pools are capped (~96 → ~56 threads per daemon).
+- Event export and the HTTP/MCP daemon-warm path route through the EventStore seam.
+- Metrics UI: an empty phase filter means ALL phases, not none.
+- `testerkit init` merges an existing `.vscode/settings.json` instead of skipping it.
+- The AI read surface is split into the `testerkit-data` skill, with a hard rule
+  against reading raw parquet.
+
+## [0.3.1] - 2026-07-06
+
+Out-of-the-box: a fresh `testerkit init --starter` runs, surfaces live and finished runs
+across every reader, and enables the VS Code Test Explorer with no setup.
+
+### Added
+
+- Live in-flight runs: the runs materializer launches with the pytest session, so a run materializes without a reader. `testerkit runs`, the HTTP API, and MCP tools show in-flight runs as `RUNNING`.
+- `testerkit init` enables the VS Code pytest Test Explorer in the generated `.vscode/settings.json`.
+- `testerkit init` gains `--no-input` / `--no-ai` for headless scaffolding.
+
+### Changed
+
+- Starter is one vectorized test now: a single inline `testerkit_sweeps` + `testerkit_limits(characteristic=…)` test (`test_output_voltage[3.3]/[5.0]/[5.5]`), no sidecar, spec-driven limit, swept `vin` recorded as an input.
+- Starter `testerkit.yaml` drops internal tuning knobs it never sets.
+
+### Fixed
+
+- `testerkit init` no longer hangs in Codespaces/CI (skips the AI-setup prompt under `--no-input`/`--no-ai`/`CI`/`CODESPACES`).
+- `testerkit init --starter` instrument assets use `id == role` so calibration joins the station.
+- `testerkit validate` detects file types by structural shape — bare `testerkit validate` on a fresh starter used to fail 3 of 6 files.
+- Per-test route cleanup no longer requests a fixture during teardown (`PytestRemovedIn10Warning`).
+- `testerkit-tests` skill: `observe` is for output evidence, `stream` its live sibling, never an input.
+- Stale `psu.yaml`/`dmm.yaml` asset names in `docs/tutorial/quickstart.md`.
+
+## [0.3.0] - 2026-07-06
+
+Execution-grain and schema-versioning release. Steps and vectors get a clean
+at-rest grain — a step carries its own measurements; vectors exist only as
+condition points (sweeps / inner loops). The at-rest schema starts a distinct
+pre-1.0 line (baseline `0.1`) with read-time version dispatch, `slot` becomes
+`site` throughout, and instruments gain per-step reservations. (The analytics
+suite moves to 0.4.0.)
+
+Pre-1.0: this release rewrites the at-rest schema, and the schema stays on a
+0.x line — each minor is a breaking epoch that battle-tests the version
+apparatus, so 1.0 is earned later, not frozen now. Regenerate `data/` from
+fresh `0.1` artifacts; older parquet is read via version dispatch or quarantined.
+
+### Changed
+
+- **BREAKING** `slot` → `site` everywhere: 0-based `site_index` (always present, default 0) and frozen `site_name`; CLI flags and STDF `SITE_NUM` follow.
+- **BREAKING** step executions de-fused — one row per execution, retries counted as occurrences instead of overwriting the prior attempt.
+- **BREAKING** `uut_serial` → `uut_serial_number` at rest and in the API.
+- **BREAKING** measurement storage reshaped to the step/vector grain — a step carries its own measurements, and per-row values move out of wide per-field columns into three nested `LIST<STRUCT>` columns: `inputs`, `outputs`, `measurements` (one struct per value). Downstream readers of the old flat columns must read the nested columns or query through `MeasurementsQuery`.
+- **BREAKING** dynamic input/output fields drop their `in_`/`out_` name prefixes at rest; discovery is role-based (EAV), so a user's unprefixed field names no longer disrupt the query surface. Reference a field by `(role, name)` (`FieldRef`) instead of a prefixed column name.
+- **BREAKING** at-rest schema reset to a `0.1` baseline — a distinct pre-1.0 line, not frozen; `parent_path` dropped (derived from `step_path`).
+- Timestamps are UTC at every server boundary; clients translate at their own edge (UI / CLI / MCP).
+
+### Added
+
+- At-rest schema versioning: a `0.1` baseline registry, whitelist-dispatch readers at all four store boundaries, and an opt-in forward-migrate sink. Newer-stamped files are deferred (a newer daemon re-reads them); unreadable ones are quarantined — never a hard crash.
+- Instrument reservations: re-entrant, timeout-aware resource locks, per-step reserve/release auto-wrap, step-duration server leases, and `instrument.reserved` / `instrument.released` events. Per-step/vector instrument sets are recorded at rest.
+- AI-skills reimagined: 11 focused Agent Skills (`testerkit-<domain>/SKILL.md`) installed per-tool by `testerkit setup` (Claude Code/Codex/Cursor/Copilot, native); a new `testerkit docs show` CLI streaming the shipped docs; removal of the old `testerkit refs` CLI / workflows / command stubs.
+
+### Fixed
+
+- Parts page shows run counts and "Observed" rows — observation now keys on the hardware `uut_part_number`, not the config-slug `part_id`.
+- Run-detail measurement counts (Overview + Steps tabs) under the reshaped grain.
+- Channel-detail live refresh under the default `testerkit serve` (the channels Flight→UI bridge was previously wired only on `--reload`).
+- The runs daemon self-heals a corrupt `_index.duckdb` at boot (rebuild from parquet) instead of crash-looping to a blank UI.
+- `/explore` defaults its X axis to a per-measurement occurrence `index` and centers a single-valued axis, so non-swept measurements plot instead of showing an empty chart.
+- `/explore` no longer emits Quasar "Anchor: target not found" console errors.
+
+## [0.2.1] - 2026-06-26
+
+Improved documentation: a corpus-wide accuracy pass across the reference,
+how-to, and operator-UI docs, plus the bug fixes it surfaced.
+
+### Fixed
+
+- Catalog variant `bands:` now append to the base signal instead of replacing it.
+- `data_dir` no longer falls back to a hardcoded `./results`; it resolves through `testerkit.yaml`.
+- Events filter and Channels list subscribe to `channel.started` (the retired `instrument.read` returned nothing).
+- Grafana measurement views and dashboards work against the schema-2.0 layout.
+- MCP and skill prompts no longer use a retired event type or invalid sidecar YAML.
+- `testerkit setup copilot` writes installer-agnostic `testerkit mcp serve` to `.vscode/mcp.json`.
+
+### Changed
+
+- Measurements (`/explore`) Part/Station/Phase filters use the operator-facing URL keys `part`/`station`/`phase`, matching Metrics. Old `?part_id=` URLs no longer apply.
+
+## [0.2.0] - 2026-06-22
+
+Data-architecture release. A fourth store (FileStore) joins Runs,
+Events, Channels for blobs / waveforms / streaming captures. Three
+test-author verbs — `observe`, `verify`, `stream` — replace ad-hoc
+measurement recording with a typed, routeable surface. The operator UI
+gains entity-observed-view across inventory pages, two new pages
+(UUTs, Profiles), and an AST-driven Tests rewrite.
+
+A follow-on pass deepens the execution data model (vector-grained,
+chronological parquet + EAV), the programmatic read surface (query
+measurements by role + name; consistent, typed query/store interfaces
+with fool-proof lifecycles), and the quality metrics (Ppk, RTY,
+DPMO/DPPM).
+
+### Added
+
+- **FileStore** at `testerkit.data.files` — session-keyed artifact store
+  with `files.write(name, value)` returning a `file://` URI. Typed
+  serializer registry (bytes, Pydantic, ndarray, Waveform, PIL.Image,
+  DataFrame, Arrow IPC) with `register_serializer(...)` for custom
+  types and a pickle fallback that warns.
+- **Streaming sink** — `files.stream(name, format=...)` for multi-chunk
+  captures. Four built-in formats: `raw`, `jsonl`, `tdms`, `h5`.
+- **Three test-author verbs.** `observe` records without judgment and
+  routes by value shape (scalars/arrays → ChannelStore, blobs →
+  FileStore). `verify` records a scalar measurement and judges against
+  a limit (non-scalar raises pointing at `observe`). `stream` explicitly
+  routes a sample to ChannelStore. All three available as Context
+  methods + bare pytest fixtures.
+- **Channel start/end events** — `ChannelStarted` / `ChannelEnded`
+  bracket every channel session.
+- **File start/end events** — `FileStarted` / `FileEnded` bracket
+  every streaming sink session. Live consumers range-read the file
+  directly via the path in `FileStarted`.
+- **Typed event payload columns** — 22 IDs and names (channel_id,
+  uut_serial, role, outcome, etc.) promoted from JSON payload to
+  typed DuckDB columns, enabling WHERE pushdown. Measured: 2.74×
+  speedup on `outcome=failed` filter over 10k events.
+- **Live waveform plot** on `/channels/{id}` updates push-style as
+  samples arrive.
+- **`XYData` model** for paired-array data (IV curves, eye diagrams,
+  S-parameter sweeps).
+- **Entity-observed-view across operator UI** — stations, parts,
+  fixtures, instruments list pages merge YAML-configured + observed-
+  in-runs entities with a Configured/Observed chip and filter.
+- **New `/uuts` page** — one row per distinct UUT serial in run history.
+- **New `/profiles` list + detail pages** — profile registry with
+  extends-chain rendering and resolved YAML view.
+- **Rewritten `/tests` page** — AST-driven file-level layout with
+  per-test panels, run history, and an "Observed in history" section
+  for orphaned step paths. Detail page at `/tests/{path}` with Code +
+  Sidecar YAML tabs.
+- **`/files` operator page** at the DATA STORES nav — list every
+  FileStore artifact with mime / size / session filters. Per-artifact
+  detail page with mime-switched viewers (image, JSON pretty-print,
+  JSONL table, CSV table, NPZ Waveform chart, NPY stats, hex
+  download). `?download=1` forces Content-Disposition save.
+- **`testerkit_files` MCP tool + `GET /api/files/catalog`** — list
+  FileStore artifacts (uri / session / run filters, newest-first) from
+  agents and HTTP, mirroring the existing `/files` byte-server.
+- **Mock noise spec** — a station `mock_config` value shaped
+  `{nominal, sigma}` returns a fresh `random.gauss(nominal, sigma)` draw
+  each call, so mock measurements vary run-to-run (real distributions /
+  Cpk / yield) instead of one repeated value.
+- **Spec-limit overlay on `/explore`** — when the Measurement filter is
+  narrowed to one measurement and Y is `measurement_value`, scatter /
+  line charts overlay that measurement's low/high limits from its most
+  recent run as black dashed lines that track the X axis (a step band
+  when the limit is condition-indexed, flat when constant).
+- **`/channels` list filters** — Channel ID contains / Type /
+  Instrument / Since-Until, URL-mirrored. Live-poll + in-place row
+  mutation pattern preserved.
+- **`/channels/{id}` chart groups by session** — scalar channels with
+  samples from multiple runs render one series per session, distinct
+  color per session, legend (scrolling, under the plot) labelled
+  `<uut_serial> · <YYYY-MM-DD HH:MM:SS>`. A **Time | Index** x-axis
+  toggle (URL-shared via `?x_mode=`) overlays sessions by sample index
+  for shape comparison. An activity-driven badge reads `● live` while
+  samples arrive, `○ idle` otherwise. Single-session and waveform
+  overlays unchanged.
+- **`/results/{run_id}` "View this run's" card** — Events, Channels,
+  and Files deep-link buttons all carry the run's session into the
+  target page (URL-only scoping via `session_filter_banner`).
+- **`/launch?test_profile=<name>`** — query param now wires through
+  to `LaunchRequest.test_profile` and `--test-profile=` on the pytest
+  cmdline. New "Profile" dropdown on the launch form.
+- **User-facing API surface re-organization.** 22 names promoted
+  across `testerkit` top-level + new `testerkit.queries` submodule + new
+  `testerkit.ui` helpers:
+  - `from testerkit import connect, observe, verify, stream, Mock, Waveform, XYData, Outcome, TesterKitClient`
+  - `from testerkit.queries import RunsQuery, StepsQuery, MeasurementsQuery, EventStore`
+  - `from testerkit.ui import page_layout, data_table, subscribe, channel_data, bind_channel_store, ...`
+
+  Deep paths still work; docs / examples / tutorials swept (47
+  files) to use the shallow paths. Verbs (`observe` / `verify` /
+  `stream`) are now importable functions in addition to the pytest
+  fixture form.
+- **Examples 08–11** ship four end-to-end demos of the data
+  architecture: waveform evidence (`observe(Waveform)` + `verify`
+  derived scalars), continuous monitoring (`channels.stream` + live
+  UI), artifacts + byte streams (`observe` PIL/bytes/Pydantic +
+  `files.stream`), querying data (consumer-side via
+  `RunsQuery` / `MeasurementsQuery` / `EventStore`). New tutorial
+  steps 11–12 + four how-to pages teach the pattern.
+- **Query measurements by role + name.** `MeasurementsQuery` selects via
+  `FieldRef` / `FieldRole` (e.g. `parametric(y="v_rail")`),
+  scoping to a measurement by role + name instead of the fused wide columns.
+- **RTY + DPMO/DPPM on the yield tab.** Rolled throughput yield
+  (`∏` per-step first-pass rate), defects-per-million-opportunities
+  (opportunity = a measurement, matching the failure-pareto), and
+  defective-parts-per-million, surfaced as headline cards + on the CLI /
+  MCP / HTTP yield summary. A pooled `MeasurementsQuery.yield_overall`
+  computes the headline numbers over the whole filtered set (distinct
+  serials once; RTY as the pooled per-step product).
+- **`format_number()`** shared UI formatter — `g`-format numeric display
+  that strips IEEE float-repr noise (e.g. limits `0.04 – 0.06`, not
+  `…0600000000000005`).
+- **`scripts/seed-demo-data.py`** — official maintained script that
+  generates a representative analytics dataset (multiple parts / serials /
+  stations, realistic failures + measurement spread, backdated over a
+  multi-day window) so the metrics screenshots and dashboards show real data.
+- **Optional-close store contract + lean `Store` Protocol.** `RunStore` /
+  `EventStore` / `ChannelStore` / `FileStore` are construct-and-reuse;
+  `with` / `close()` are optional, and `EventStore` + `ChannelStore` carry
+  `weakref.finalize` nets so a forgotten `close()` can't leak the
+  in-process resources.
+- **Vector-grained execution model.** `VectorStarted` / `VectorEnded`
+  events + retry-aware step events; measurements nested under the vector
+  in a chronological-telling parquet with an EAV projection. Observation
+  pinning records `uut_pin` on output lanes.
+
+### Changed
+
+- **`testerkit benchmark` reports concurrency per store, measured honestly.**
+  The write-throughput sweep (1 → 2 → 4 writers) now runs for **events,
+  channels, files, and runs** — each with its own scaling curve and
+  per-writer efficiency — instead of a single runs-only sweep collapsed into
+  one headline factor (dropped). Two measurement fixes: the round wall is now
+  the true overlapped span (`max(end) − min(start)` across workers, not the
+  slowest worker's self-timed loop, which overstated throughput); and the
+  runs workers no longer perform a synchronous `notify_new_run` the real
+  `save_test_run` path never does (that phantom ACK was artificially
+  serializing the runs sweep). New `report.json` key `concurrency_by_store`
+  (replaces `concurrency_sweep`).
+- **Metrics are per-phase now.** The quality dashboards (Yield / Pareto /
+  Cpk / Retest) default to **`phase = production`** and the **last 30 days**
+  instead of "all phases except development, all time" — production is the
+  only phase where FPY / Cpk / pareto are meaningful (development is mock /
+  dirty-git data; characterization deliberately drives out-of-spec). The
+  Phase filter switches it. The empty state names the cause ("No production
+  runs in the window — change the Phase filter or widen the date range")
+  rather than the misleading "No measurements yet."
+- **BREAKING — `product` → `part` and `dut` → `uut` rename.** The
+  type/definition entity is now **Part** and the physical instance is
+  now **UUT** (unit under test), aligning with ASAM AoPart / STDF /
+  ATML terminology and generalizing beyond electronics. Renamed
+  across the YAML schema (`products/` → `parts/`), pytest fixtures,
+  event/parquet fields (`uut_serial`, `uut_part_number`), parquet
+  columns, operator-UI routes (`/parts`, `/uuts`), and the HTTP / MCP
+  query surfaces. Pre-1.0, there is no compatibility shim: existing
+  `product_*` / `dut_*` YAML and parquet must be regenerated or
+  migrated.
+- **Operator-readable session labels everywhere.** Pages that displayed
+  session UUID prefixes now resolve to `<uut_serial> · <YYYY-MM-DD
+  HH:MM:SS>` via a shared lookup helper (`/channels` data tab,
+  `/files` detail Session field, the session filter banner on
+  `/events` / `/channels` / `/files`). Banner distinguishes the
+  filter-active state (blue) from the session-not-found state (amber)
+  so stale bookmarks have explicit copy.
+- **Session URL param standardized to `?session_id=`** across
+  `/channels`, `/channels/{id}`, `/files`, `/events`, `/results/{run_id}`
+  deep-links. Bookmarks using the prior `?session=` form on `/channels`
+  no longer carry; deep-link from a fresh `/results/{run_id}` to
+  rebuild.
+- **`namespace=` parameter parity.** `files.write` and `files.stream`
+  now accept `namespace=` matching the existing `channels.write` /
+  `channels.stream` / `observe` / `verify` / `stream` surfaces. An
+  artifact recorded via `observe(name, value, namespace="psu")` and
+  one recorded via `files.write(name, value, namespace="psu")` land at
+  the same effective name.
+- **`FileStore` serves artifact bytes by URI** via `read(uri)` /
+  `read_range(uri, ...)` / `open_input(uri)`. The UI service,
+  materializer, `/files-static` route, and HTTP API read through these.
+- Channel detail Sequences tab renamed to Capabilities (the tab's
+  content was always station capabilities; sequences are deferred).
+- ChannelStore schema: `timestamp` → `received_at` (store-side, always
+  present) + new nullable `sampled_at` (hardware-side); `samples` →
+  `value` (unify scalar/array payload column); `properties`/`attrs` →
+  `attributes`; typed leaf-types (`bool`, `int`, `float`, `str` for
+  scalar and array shapes).
+- Materializer auto-promotion: observation-only vectors now produce
+  `DONE` rows in the parquet measurement table instead of disappearing
+  from the analytical view.
+- Two new operator-UI reference pages (`uuts.md`, `profiles.md`) bring
+  the total to 18. Four reference pages updated for the chip + filter.
+  Tests reference page rewritten for the AST layout.
+- **BREAKING — `Cpk` / `Cp` → `Ppk` / `Pp`.** The capability metric computes
+  `STDDEV_SAMP` (overall, long-term σ) — which is Ppk/Pp's basis, not
+  within-subgroup Cpk/Cp — so it was renamed for honesty across
+  `MeasurementsQuery`, the CLI command, the MCP action, the `/metrics/ppk`
+  route, the operator-UI tab, and the docs. True Cpk (within-subgroup /
+  I-MR σ) is deferred to v0.3.0.
+- **BREAKING — measurements queried by role + name.** The fused
+  `out_<name>` / `in_<name>` wide columns are dropped;
+  `RunStore.get_measurements` and the run-detail read surface return
+  role-split inputs / outputs.
+- **BREAKING — channel schema: `data_type` → `value_type`** (cross-store
+  datatype-name consistency) and channel `offset` → `sample_offset`. A
+  channel's `unit` is unified with the channel descriptor and fails loud
+  on conflict (matching `stream`).
+- **Scalar `units` → `unit`** across the recording verbs and models
+  (collections stay plural); `unit=` is accepted on every recording verb.
+- **Consistent query/store interfaces.** Uniform query method names + a
+  shared `dynamic_attrs` decoder; typed `Row` returns + `ColumnSchema`
+  describe-columns; `FileStore`'s `data_dir` is now the private
+  keyword-only `_data_dir` (an infrastructure path, not user-facing).
+- `execution/logger.py` → `run_scope.py` (and the `logger` fixture →
+  `_run_scope`); `StreamCheckpoint` split into `ChannelCheckpoint` +
+  `FileCheckpoint`; store-lifecycle events renamed for grammatical
+  consistency.
+- The ~3k-line `cli.py` monolith split into a `cli/` package; the System
+  Designer page marked experimental. `run_id` is now in the per-run
+  parquet filename (prevents silent overwrite).
+
+### Removed
+
+- `InstrumentRead` event class — per-sample events at DAQ rates flooded
+  the EventStore. Sample data lives in ChannelStore; start/end events
+  (`ChannelStarted` / `ChannelEnded`) replace it.
+- `StreamFrameIndex` event class — same flood problem at chunk rate.
+  Live consumers read the file directly.
+- Dormant measurement-ref struct slot on the run schema (unused).
+
+### Fixed
+
+- **Live-channel UI panels deliver on the event loop.** Channel-data
+  callbacks fired on the Flight reader thread, mutating NiceGUI elements
+  off the loop; they now marshal through the UI loop (same contract the
+  event path already used).
+- `/channels/{id}` 500 error on pure-scalar channels.
+- **`/explore` default axes pick a real parameter.** A measurement-
+  scoped URL with no axis params fell through to the first schema
+  column — often an id like `characteristic_id` for X and a stimulus
+  input for Y. X now prefers a swept `in_*` input (then `vector_index`
+  / time, never an `*_id`); Y prefers `measurement_value`.
+- `Context.stream` and `channels.write` did not emit `ChannelStarted`
+  (only the observer path did). All writer paths now emit it.
+- `/live/{run_id}` Streams panel showed streams from all runs;
+  subscription now scopes by `run_id`.
+- **`observe()` URIs reach parquet `out_*` columns.** Before this
+  fix, `context.observe("uut_photo", img)` wrote the file to
+  FileStore but the URI lived only on `Context._observations` —
+  `logger.log_measurement` projected `out_*` from
+  `vector.observations` (empty), so the operator UI's Measurements
+  tab showed no artifacts. Now `observe` mirrors to the active
+  vector at write time.
+- **Warm-query perf gates stabilized.** Two single-shot perf timers
+  flaked under suite load (~30% of full runs). Hard caps (100ms / 200ms)
+  unchanged; sampling now takes the
+  min over 11 calls so transient spikes don't trip the gate.
+- **`StepBuilder` propagates `PASSED` step outcomes** on the default-vector
+  (`step.measure()`) path. Previously only `FAILED` propagated, so a passing
+  step written via the catch-all `TesterKitClient` results API ended
+  `outcome=None` — excluded from step counts and leaving the run outcome
+  `None`. Now consistent with the explicit `step.vector()` path.
+- **Yield headline cards are pooled.** They previously combined the
+  per-(part × station × period) rows, which broke RTY (multiplied as a
+  product across groups) and Final Yield (double-counted serials across
+  groups, so Final < FPY). `yield_overall` computes them over the whole
+  filtered set instead.
+- **Daemon write-path resilience.** A daemon killed mid-write reacquires a
+  fresh one and resends the un-acked batches (idempotent via
+  `ON CONFLICT (id) DO NOTHING`); subscribers resume from their cursor; a
+  spawn that times out fails fast and surfaces the crash reason.
+
+### Deferred to v0.3.0
+
+- **Local shared-memory transport** (item 22) — PoC measured 2×
+  latency over Flight loopback, not the 3–10× estimated. Revisit on
+  symptoms (UI lag traced to transport; >10 kHz capture saturation).
+- **Consumer SDK `testerkit.live`** (item 20) — store APIs are
+  available; higher-level consumer surface not yet designed.
+- **Hardware video encoder formats** (item 23) — `mp4` / `wav` /
+  `flac` handlers on top of the streaming-sink machinery.
+- **Channels + files as test INPUTS** — current verbs are producer-only;
+  a first-class input surface (`channels.read` / `files.read`) lands
+  in v0.3.0.
+- **Analytics metrics release** — true Cpk (within-subgroup / I-MR σ,
+  distinct from the overall-σ Ppk shipped here), per-measurement SPC
+  control charts (I-MR / X̄-R + Western Electric rules), yield cross-tab by
+  station / fixture / operator, and a generic `pareto(by=measure)`. Roadmap:
+  `docs/_internal/explorations/0.3.0-analytics-metrics.md`.
+
+## [0.1.3] - 2026-05-24
+
+Documentation-heavy release. Full per-screen reference for the operator
+UI plus a Diátaxis × topic reorg that lands every doc page in a
+predictable `<quadrant>/<category>/<topic>.md` cell, with cross-quadrant
+"See also" navigation between them.
+
+### Added
+
+- **16 per-screen reference pages** under `docs/reference/operator-ui/`
+  — every NiceGUI page (Dashboard, Launch Test, Live monitor, Results
+  list/detail, Metrics, Measurements, Events, Channels list/detail,
+  System Designer, Stations, Parts, Fixtures, Instruments, Tests)
+  documented from the running source. Each carries a cropped
+  testid-anchored screenshot.
+- **Cropped UI screenshots** via new `scripts/regenerate-ui-screenshots.py`
+  (Playwright + headless `testerkit serve`) — manifest-driven, PNGs commit
+  into `docs/_assets/operator-ui/`.
+- **Tour bridge** at `docs/how-to/overview/operator-ui-tour.md` —
+  orientation map of all 14 sidebar entries.
+- **Diagnostic how-tos**: `find-flaky-tests`, `compare-runs`,
+  `export-results`, `operator-prompts`, plus two MCP-driven recipes
+  (`mcp-query-runs`, `mcp-debug-failures`).
+- **Grafana integration** reference at `docs/integration/data/grafana.md`
+  (pgwire data source + ten shipped dashboards).
+- **Live API explorer** subsection in `reference/runtime/api.md`
+  covering Swagger UI / ReDoc / OpenAPI JSON.
+- **Tutorial step 10** expanded with Results + Metrics walkthroughs.
+- Pre-commit hook `screenshot-drift-reminder` — non-blocking nudge to
+  rerun the screenshot script when a UI file with a manifest-tracked
+  testid changes.
+
+### Changed
+
+- **Docs reorg — Diátaxis × topic matrix.** 64 file moves into
+  per-category subdirectories (`overview/` / `configuration/` /
+  `execution/` / `data/` / etc.) so the matrix axis is consistent across
+  quadrants. Same path tail across quadrants gives natural cross-links:
+  `concepts/configuration/fixtures.md` ↔
+  `how-to/configuration/configuring-stations.md` ↔
+  `reference/configuration.md`. Filename prefixes (`testerkit-`,
+  `catalog-`, `why-`) dropped — directory carries the namespace. Three
+  former "why-" concept pages rewritten to read as concept references,
+  not blog posts.
+- **In-app docs viewer** now supports nested subdirectory pages: route
+  changed to `{page:path}`, `_parse_section_outline` switched to
+  `rglob`, sidebar groups render as accordion (current page's group
+  auto-expands).
+- **Operator-UI table cleanup**: dropped run UUIDs from operator-facing
+  tables (operators identify runs by UUT serial + start time, not
+  UUID). Events page Session filter changed from free-text input to
+  autocomplete dropdown labelled `<timestamp> • <client>` (pytest /
+  jupyter / etc.). Live monitor's Run ID kept but de-emphasized (still
+  copyable for URL bookmarking).
+- `docs/_assets/` now bundled into the wheel (was missing — operator-UI
+  screenshots wouldn't render on `pip install`ed copies).
+- Generator path table (`scripts/generate_reference_docs.py`) updated
+  for the 5 moved generated reference pages; pre-commit drift-hook
+  regex follows.
+
+### Fixed
+
+- **Test race**: a run-status test helper polled for `RunStarted` then
+  read partially-materialised steps; it now waits for the run to
+  finalise (`ended_at IS NOT NULL`) before reading.
+- **System Designer auto-save**: edits to a fixture silently failed to
+  save — the save raised an error that was swallowed, so changes weren't
+  persisted. Auto-save now writes the fixture correctly.
+- Docs viewer's `/docs/_assets/` path was being gobbled by the
+  `{page:path}` catch-all and returning HTML instead of PNGs. Mounted
+  as static files before the dynamic route.
+
+## [0.1.2] - 2026-05-19
+
+First installable PyPI release. Both 0.1.0 and 0.1.1 wheels shipped without
+`testerkit/data/` due to an over-broad `data` exclude pattern in
+`pyproject.toml`, so the bundled pytest plugin failed to import on every
+fresh install; those releases are yanked.
+
+### Added
+
+- `verify(...)` and `logger.measure(...)` accept a plain dict for `limit=`
+  (coerced via `Limit.model_validate`). Tutorials and examples now use the
+  dict form; `from testerkit import Limit` stays available for the model object.
+- `verify_requires_limit: bool | None` on `ProfileConfig` — set to `False`
+  on a characterization profile to route `verify()` to record-only
+  semantics when no limit resolves (instead of `MissingLimitError`).
+- `testerkit refs list` / `testerkit refs show <topic>` — stream curated reference
+  docs (`tiers`, `verify`, `mocks`, `profiles`) to stdout. CLAUDE.md
+  templates now point agents at this CLI instead of baking absolute paths.
+
+### Fixed
+
+- Packaging: scoped the `data` exclude pattern in `pyproject.toml` to
+  `/data` (top-level only) so `src/testerkit/data/` ships in the wheel.
+- Run outcome stamping is now retry-aware. A test that errors on attempt 1
+  and passes on the `testerkit_retry` retry stamps the RUN as `passed`
+  (matching pytest-rerunfailures, STDF MIR.RTST_COD, and Jenkins flaky-
+  test-handler conventions). The errored attempt's step row stays in
+  the run for retest / flake analysis.
+
+## [0.1.0] - 2026-04-15
+
+Initial public release on PyPI as `testerkit`.
+
+### Added
+
+- Pytest-native hardware tests — plain `def test_*` functions with
+  fixtures and markers for vector expansion, limit checking, measurement
+  recording, retries, and mock injection
+- Station / fixture / part / sequence YAML configuration, loaded through a
+  single store layer with Pydantic validation
+- Instrument fixtures resolved from station config (no `conftest.py`
+  boilerplate required)
+- `--mock-instruments` mode for hardware-free development
+- Parquet result storage with per-step instrument traceability
+  (serial, cal due date, firmware)
+- DuckDB-backed analytics layer over the Parquet silver/gold layout
+- Operator UI (`testerkit serve`) built on NiceGUI
+- FastAPI HTTP API and MCP server, with parity between the two
+- Capability matching (`testerkit_match`) against an instrument catalog
+- CLI: `testerkit init`, `discover`, `station init`, `new-test`, `serve`, `runs`,
+  `show`, `instrument list`, `mcp serve`, `setup`
+- Optional extras for output formats (`stdf`, `hdf5`, `tdms`, `mdf4`),
+  transports (`s3`, `gcs`, `azure`, `sftp`), and integrations (`pymeasure`,
+  `ni`, `lxi`, `grafana`, `pdf`, `sbom`)
+
+[Unreleased]: https://github.com/pragmatest-dev/testerkit/compare/v0.2.0...HEAD
+[0.2.0]: https://github.com/pragmatest-dev/testerkit/compare/v0.1.3...v0.2.0
+[0.1.3]: https://github.com/pragmatest-dev/testerkit/compare/v0.1.2...v0.1.3
+[0.1.2]: https://github.com/pragmatest-dev/testerkit/releases/tag/v0.1.2
+[0.1.1]: https://github.com/pragmatest-dev/testerkit/releases/tag/v0.1.1
+[0.1.0]: https://github.com/pragmatest-dev/testerkit/releases/tag/v0.1.0