@woosh/meep-engine 3.1.0 → 3.2.0

package/src/generation/REVIEW_01_ACTION_PLAN.md

@@ -1,628 +1,628 @@
-# Generation Package — Review 01 Action Plan
-
-Source: architectural + correctness review of `meep/src/generation`, 2026-06-11.
-Paths below are relative to this directory unless noted.
-
-## Scope rulings (from review discussion)
-
-These override the raw review findings:
-
-1. **No string-keyed factories.** One-class-per-operation stays. We accept the
-   verbosity of `CellFilterAdd` etc. in exchange for typed, greppable, JSDoc-able
-   code. Boilerplate reduction must come from better *base classes*, never from
-   `X.of('add', fn)`-style magic.
-2. **This is a library.** "Nothing imports it" is not an argument for removal.
-   Unreferenced classes are public surface: bugs in them get **fixed**, not
-   deleted. Removal is only on the table for code that *cannot work for any
-   caller* (non-functional stubs), and that is a per-item decision.
-3. **`prototype*.js` files are internal prototyping tools**, never published.
-   Out of scope for this plan.
-4. House rules that shape the fixes: correctness first (no bandaids), black-box
-   tests with real objects (no mocks, no call-count assertions), Jest runs from
-   the `moh` root with the root `jest.conf.json`, asserts are free in prod,
-   module-static pools are accepted style (guard them, don't remove them),
-   prefer required parameters over null-means-default sentinels.
-
-**Output-stability flag:** items marked `[output-changing]` alter generated
-levels for a fixed seed. Each such fix must be verified visually (see
-*Verification protocol* at the bottom) and batched consciously — they are the
-reason this plan is phased.
-
----
-
-## Phase 1 — Surgical bug fixes
-
-Local, mechanically obvious fixes. Each lands with a black-box spec.
-
-- [x] **1.1 `MarkerNodeMatcherAnd.match` implements OR.**
-  `markers/matcher/MarkerNodeMatcherAnd.js`
-  Fix: `if (!this.left.match(node)) return false; return this.right.match(node);`
-  Test: spec with two `MarkerNodeMatcherByType`/`ContainsTag` matchers over a real
-  `MarkerNode`; truth-table all four combinations. Do the same table for `Or` and
-  `Not` while in there — the family has zero coverage.
-  `[output-changing]` — restores intended semantics in
-  `app/src/generator/generator/objectives/GENERATE_BOSS_OBJECTIVE.js` and
-  `generateMusic.js`. While fixing, read both call sites and confirm AND is what
-  they want (it is what they wrote).
-
-- [x] **1.2 Road relaxation typo.**
-  `grid/generation/road/GridTaskGenerateRoads.js:287`
-  Fix: `distances[neighbour_index] = distances;` → `= distance;`
-  Test: covered by the end-to-end roads test in 1.3 (the branch is hard to force
-  in isolation; the e2e test plus an `assert.isNumber` in the loop is enough).
-
-- [x] **1.3 Road groups self-connect and are then skipped.**
-  `grid/generation/road/GridTaskGenerateRoads.js` (node-encounter branch, ~line 336)
-  Fix: skip own-group nodes the same way the path-encounter branch skips own-group
-  paths: `if (targetGroupId === groupId) continue;` before the
-  `directGroupConnectionExists` check.
-  Test (e2e, black-box): small `GridData`, all-traversable layer, two marker
-  groups of **two adjacent connectors each**, placed apart; run the task to
-  completion; assert a `RoadConnection` exists whose endpoints belong to
-  *different* groups, and that every group is reachable in the connection graph.
-  Note: requires the matcher/action parameters from 5.1, or temporarily importing
-  the sample matchers in the spec (acceptable for the test, not for src).
-  Also in this file: delete the leftover `console.log(path)` in `tDrawPaths`, and
-  the constructed-but-unused `Graph` instance.
-
-- [x] **1.4 `GridTaskExecuteRuleTimes` never terminates when the pattern can't match.**
-  `grid/generation/GridTaskExecuteRuleTimes.js`
-  Fix: `closed.set(index, true)` after each attempted cell (match or not), and
-  make the exhaustion check count closed cells (`closedCount >= gridSize`)
-  instead of the linear-probe counter `i >= endIndex`.
-  Semantics note: this closes cells whose match could become true later as the
-  grid mutates. That is the original design intent (the `closed` set exists);
-  document it in the class doc.
-  Test: rule whose pattern never matches → task ends with failure within one
-  pass; rule that can match exactly `k < count` cells → terminates, `k`
-  placements. Used live in `AlienGridGenerator` / boss objective — confirm the
-  failure signal (`EndFailure`) is acceptable to those flows or whether
-  exhaustion should be `EndSuccess` with fewer placements. **Decision:** failure
-  vs. partial success on exhaustion.
-
-- [x] **1.6 `MarkerNodeTransformerYRotateByFilter` Euler convention mismatch.**
-  `markers/transform/MarkerNodeTransformerYRotateByFilter.js`
-  Fix: `toEulerAnglesXYZ` → `toEulerAnglesYXZ`, matching the recompose call and
-  the sibling `…YRotateByFilterGradient`.
-  Test: node with a known non-trivial prior YXZ rotation (x ≠ 0), constant
-  filter; assert resulting quaternion equals `fromEulerAnglesYXZ(x, expectedY, z)`.
-  `[output-changing]` for rotated decor in `theme-1/theme-2 injectDecor`.
-
-- [x] **1.7 `ThemeEngine.applyNodes` searches a stale-length influence array.**
-  `theme/ThemeEngine.js` (applyNodes)
-  Fix: pass explicit bounds to `binarySearchLowIndex(themeInfluence, t, cmp, 0, matchingThemeCount - 1)`
-  (mirroring the weighted selects). Handle `influenceSum === 0`: pick uniformly
-  among the matching themes using the same `random` stream (deterministic), since
-  "all themes at mask-distance 0" is a legitimate boundary state.
-  Test: grid with two overlapping `AreaTheme` masks; a marker matching 3 themes
-  processed immediately before one matching 1 theme; assert the second node is
-  processed by its only matching theme. Use a test-local recording
-  `MarkerNodeAction` subclass (a real action, observable behaviour — not a mock).
-  `[output-changing]` at theme boundaries.
-
-- [x] **1.8 `MarkerNodeActionPaintTerrain`: clamp the splat region.**
-  `markers/actions/terrain/MarkerNodeActionPaintTerrain.js`
-  Fix (this phase, pure bug): clamp `x0/x1/y0/y1` to `[0, splat_resolution-1]`
-  so edge markers stop wrapping writes into the opposite row. Make the X/Y
-  region rounding symmetric (`floor(v + 0.5)` both axes).
-  The deeper grid↔texel mapping question (the `+ splat_to_grid` "cell center"
-  shift) is deferred to Phase 3 — leave a `TODO(review-01 §3)` comment.
-  Test: marker at grid (0,0) with radius reaching off-map; assert no texel in the
-  rightmost column / other rows changes. `[output-changing]` within one texel of
-  map edges only.
-
-- [x] **1.9 `ContinuousGridCellActionWriteObstacle` writes global coords.**
-  `grid/actions/ContinuousGridCellActionWriteObstacle.js`
-  Fix: `to.writePoint(local_x, local_y, …)`. Library surface — fix despite no
-  internal callers. Test: obstacle target with non-zero `GridPosition`; execute at
-  a global coordinate; assert the *local* cell is written.
-
-- [x] **1.10 `MarkerNode.overlaps` compares the wrong plane.**
-  `markers/MarkerNode.js`
-  Fix: use `this.position` / `other.position` (grid-space Vector2), matching
-  `GridDataNodePredicateOverlaps`. Library surface; fix, add a 3-case spec
-  (overlap / touch / disjoint).
-
-- [x] **1.11 `CellFilterLiteralBoolean` cannot execute.**
-  `filtering/boolean/CellFilterLiteralBoolean.js`
-  Fix: implement `execute() { return this.value ? 1 : 0; }` (the package-wide
-  boolean-as-number convention, same as `CellFilterCellMatcher`). Add `TRUE` /
-  `FALSE` singletons alongside `CellFilterLiteralFloat.ONE/ZERO` for symmetry.
-  Spec: both values, plus `dataType === DataType.Boolean`.
-
-- [x] **1.12 Weighted selects: total weight 0 still executes an element.**
-  `placement/action/random/weighted/CellActionSelectWeightedRandom.js`,
-  `markers/actions/probability/MarkerNodeActionSelectWeighted.js`
-  Fix: if `totalWeight <= 0`, execute nothing. (Folded into 5.2's shared helper;
-  if 5.2 is deferred, fix both copies now.)
-  Test: all-zero weight filters → no action executed (recording action).
-
-- [x] **1.13 Small fixes batch** (no individual specs needed beyond existing suites):
-  - `GridTaskGroup.catchGeneratorErrors` — include `reason` in the
-    `console.error` (currently drops the actual error).
-  - `GridTaskApplyActionToCells.from` — apply the `name` parameter; rename the
-    misleading `fitness` field to `supplier` (or `selector`) to match the API.
-  - `CellSupplierBestN` — implement `count()` (`this.elements.length`) so task
-    progress works.
-  - `GridTaskGenerateRoads` — assert integer connector positions before
-    `index = x + y * width` (fractional positions silently corrupt the search).
-  - `GridData.resize` — assert message typo (`'width'` on the height check).
-  - `CellFilterInverseLerp.from` — JSDoc for `f`/`value` is copy-pasted from Lerp.
-  - `sampler_from_filter` — assert says non-negative, doc says positive; make it
-    `greaterThan(0)`.
-  - `GridTaskConnectRooms` — move the unconditional `drawSamplerHTML` import into
-    the (commented) debug path or a lazy import so headless bundles don't pull
-    DOM helpers. Keep `debugState`/`drawDebugState` (internal tooling).
-  - `GridActionRuleSet.process` — seed the `Random` policy iterator
-    (`ArrayIteratorRandom.setSeed(seed)`) so rule order responds to the world
-    seed. `[output-changing]` for `Random`-policy rule sets.
-
-- [x] **1.14 Re-entrancy guards on module-static pools.**
-  Module-static pools are house style and stay. Add cheap in-use assert guards
-  (free in prod) where recursion through user code is possible:
-  `GridData.marker_collection` (matcher could re-query the grid),
-  `CellFilterCurvature`'s normal scratch (curvature-of-curvature),
-  `CellFilterLookupTable.sample`, `RoadConnectionNetwork.scratch_paths`.
-  Pattern: `assert.ok(!pool_in_use, 're-entrant use of <pool>')` + try/finally
-  flag toggle is NOT acceptable (silent catch rule) — use a plain set/unset; an
-  assert firing in dev is the goal.
-
----
-
-## Phase 2 — Semantic decisions, then fixes
-
-Each item changes generated output by design. Decide the intended semantics,
-fix, then verify via the sample generators (see Verification protocol).
-
-- [x] **2.1 Cellular automata: synchronous update.**
-  `automata/CaveGeneratorCellularAutomata.js`, `grid/generation/discrete/GridTaskCellularAutomata.js`
-  Current `step` mutates in place → sequential CA with scan-order bias.
-  Decision A: change `step(data, …)` to `step(src, dst, …)` and ping-pong the two
-  buffers across the 50 iterations in `GridTaskCellularAutomata` (signature
-  change to the `CellularAutomata` base — library surface, sweep subclasses).
-  Decision B (same time): out-of-bounds neighbor policy. Current clamping
-  double-counts edge cells; the conventional cave-CA choice is OOB = wall
-  (alive), which seals cave borders. Margin seeding interaction: margins start
-  dead but can currently come alive — decide whether margins should be frozen.
-  Also: write `0/1`, not `false/true`, into the `Uint8Array`.
-  Test: synchronicity — a glider-style asymmetric seed must produce the same
-  result as a reference two-buffer implementation; plus determinism (same seed →
-  identical field). `[output-changing]` — all CA-generated caves.
-
-- [x] **2.2 Fractal noise octave direction.**
-  `filtering/numeric/complex/CellFilterSimplexNoise.js` (`fractal`)
-  `scale` is a wavelength; passing `scale * frequency` makes successive octaves
-  *lower* frequency at *lower* amplitude — inverted fBm. Intended:
-  `scale / frequency`.
-  Decision: fix to conventional fBm, or keep current output and rename/document?
-  Recommendation: fix; the parameter names (`octaves`, `persistence`,
-  `lacunarity`) promise conventional fBm. Grep `fractal(` call sites first and
-  eyeball each generated surface after.
-  Test: spectral sanity — with 3 octaves, correlation length of the output must
-  *decrease* as octave count rises (or simpler: octave-2 filter alone must have
-  shorter wavelength than octave-1; assert via zero-crossing counts along a row).
-  `[output-changing]`.
-
-- [x] **2.3 Gaussian blur sigma units.**
-  `filtering/numeric/complex/CellFilterGaussianBlur.js`
-  `sigma = radius * 3` in kernel-index units makes the kernel flat (a box blur).
-  For a real Gaussian: sigma ≈ kernel half-width / 3, i.e.
-  `sigma = (samples - 1) / 6` in index units.
-  Decision: is the box-like character relied upon visually? If yes, keep and
-  rename the comment; if no, fix the constant. Recommendation: fix, since the
-  class advertises Gaussian and callers tuned radius, not shape.
-  Test: kernel center weight vs. edge weight ratio > some threshold (e.g. center
-  ≥ 3× edge for quality 3) via `build_gaussian_kernel_2d` inspection — that's
-  testing the helper's contract, not implementation internals.
-  `[output-changing]` — all blurred filters.
-
-- [x] **2.4 Heightfield normal convention (sign of gradient).**
-  `filtering/numeric/complex/CellFilterAngleToNormal.js`,
-  `filtering/numeric/complex/CellFilterCurvature.js`
-  Both build normals as `(+dX, +dY, 2)`; a true surface normal is
-  `∝ (−dX, −dY, 2)`. AngleToNormal is unaffected for the default vertical
-  reference; directional references get mirrored slopes; Curvature's sign is
-  inverted vs. its cited reference shader.
-  Fix: extract one shared `computeFilterSurfaceNormal(result, filter, grid, x, y)`
-  helper (also de-duplicates the two implementations), with the negated-gradient
-  convention, and document the axis mapping (grid x,y,height → normal x,y,z).
-  Audit the two consumers for baked-in compensations before flipping.
-  Test: synthetic ramp filter (`h = x`): normal must tilt *toward −x*; curvature
-  of a paraboloid bump must be positive at the peak (pick and document the
-  convention).
-  `[output-changing]` where non-default references / curvature thresholds are used.
-
-- [x] **2.5 `GridActionRuleSet` rotation/probability coupling.**
-  `markers/GridActionRuleSet.js`
-  Today each rotation re-rolls probability: symmetric patterns get effective
-  `1−(1−p)⁴` and rotation 0 is favored.
-  Decision: intended semantics. Recommendation: roll once per (cell × rule ×
-  pattern-offset), then scan rotations for the first match — "does this rule
-  fire here?" and "in which orientation?" become independent.
-  Test: probability 0.5 rule, rotation-symmetric matcher, large grid, fixed
-  seed: firing rate within tolerance of 0.5 (black-box statistical bound);
-  rotation histogram roughly uniform across matching rotations if that's the
-  chosen semantic. `[output-changing]`.
-
-- [x] **2.6 ThemeEngine blending math.**
-  `theme/ThemeEngine.js`, `theme/TerrainLayerRuleAggregator.js`
-  - Remove the `influence = distance / matchingThemeCount` divisor — it cancels
-    in both consumers (verified no-op; pure simplification, output-stable).
-  - `aggregator.normalize(255)` normalizes by **L2**; splat weights conventionally
-    sum to 255 (L1). Investigate what the terrain shader expects before changing.
-    If shader renormalizes, L2→L1 is output-stable; otherwise `[output-changing]`.
-  - Guard `d === 0` (no theme contributes) explicitly instead of relying on
-    `NaN → uint8 0`.
-  - `tResample` half-texel: source coords should be
-    `(x + 0.5) / splatWidth * weights.width − 0.5`; decide if the visual shift
-    is worth fixing while in here. `[output-changing]` (sub-texel).
-
-- [x] **2.7 Stateful transformers vs. estimation/test paths.**
-  `grid/generation/GridTaskDensityMarkerDistribution.js` (`estimateTapCount`),
-  `markers/GridCellActionPlaceMarkerGroup.js` (`testNodeCollisions`),
-  `markers/transform/MarkerNodeTransformerRecordUniqueRandomEnum.js`
-  `buildNode` runs transformers; estimation/collision-test paths therefore
-  *consume* from stateful transformers (UniqueRandomEnum's pool drains before
-  real placement).
-  Fixes, in order of preference:
-  1. `estimateTapCount` doesn't need nodes at all — it uses only `node.size`.
-     Compute size as `action.size * scale.sampleRandom(random)` directly; no
-     `buildNode` call. (Current transformers never alter `size`; assert that
-     invariant in `buildNode` if we rely on it — or just accept estimate drift.)
-  2. `testNodeCollisions`: grep callers; it wants post-offset positions, so it
-     can't skip transformers blindly. Split `buildNode` into the geometric phase
-     (offset/rotation/transform composition) and the transformer phase as two
-     named methods; `testNodeCollisions` uses the geometric phase. Document the
-     residual inaccuracy for position-moving transformers, or accept running
-     transformers there and document the purity requirement.
-  3. Document on `MarkerNodeTransformer`: transformers should be pure w.r.t.
-     their own state per `transform` call *or* tolerate being called for
-     discarded nodes. `RecordUniqueRandomEnum` inherently can't — note it in its
-     class doc as incompatible with estimating tasks.
-  Test: DensityMarkerDistribution with a UniqueRandomEnum transformer whose pool
-  is exactly the expected placement count → must not throw "All values used up".
-
-- [x] **2.8 Seed plumbing decisions.**
-  - `MarkerNodeEntityProcessorRandomRotation` ignores the world seed because
-    `MarkerNodeEntityProcessor.initialize(data, ecd)` has no seed parameter.
-    Add `seed` to the processor interface (sweep: ClingToTerrain, Sequence,
-    RandomRotation). Folds into Phase 4's signature unification — do it there.
-    `[output-changing]` (entity rotations start varying with seed — that's the point).
-  - `GridCellActionPlaceMarkerGroup`'s module-global `idCounter`: uniqueness is
-    all that's required today, but it breaks cross-session reproducibility of
-    `properties.groupId` values. Decision: leave (document "IDs are opaque,
-    equality-only") or derive from a per-generation counter owned by the task.
-    Recommendation: document as opaque now; revisit if save/replay needs arise.
-  - Standardize derived seeds (`seed + i`, `(s+1)*31`, `seed + 91234`, …) on one
-    small typed helper (e.g. `derive_seed(seed, salt)`) — mechanical, low risk,
-    `[output-changing]` wherever applied, so batch it with other output-changing
-    work, not piecemeal.
-
----
-
-## Phase 3 — One grid↔sampler coordinate convention
-
-The package has three conversions for the same mapping: corner-aligned
-`x/(w−1)·(sw−1)` (`GridCellActionWriteFilterToLayer`, `CellMatcherLayerBitMaskTest`,
-`ContinuousGridCellActionSetTerrainHeight`, `populateSampler2DFromCellFilter`),
-ratio `x·(sw/w)` (`AbstractCellFilterSampleGridLayer` → both sample-layer
-filters), and PaintTerrain's bespoke shift. Writes and reads through different
-conventions drift by up to a cell at the far edge whenever `resolution ≠ 1`,
-and `/(w−1)` divides by zero on 1-wide grids.
-
-- [x] **3.1 Write the mapping spec.** One short doc section (in this file or a
-  `COORDINATES.md`): grid integer coordinates are cell centers; layer texel `t`
-  covers grid `[t/r, (t+1)/r)`; conversion is the ratio form
-  `texel = grid · resolution` (consistent with `GridDataLayer.resolution` and
-  `GridData.computeScale`'s half-tile world offset). Define the out-of-bounds
-  policy per operation class: *reads* clamp or return a defined exterior value
-  (pick one), *matches* fail (no silent edge-clamping), *writes* are discarded
-  with bounds checks.
-- [x] **3.2 Implement typed helpers** (plain functions, no string keys):
-  `grid_to_texel(x, resolution)`, `texel_to_grid(t, resolution)` or equivalent —
-  small enough to inline, but one canonical definition to reference.
-- [x] **3.3 Migrate the corner-convention sites** listed above to the ratio
-  convention. Each migration gets a `resolution = 2` round-trip spec
-  (write filter → layer → read via `CellFilterSampleLayerLinear`, assert values
-  at interior and far-edge cells). `[output-changing]` for `resolution ≠ 1`
-  layers (game currently uses resolution 1 — verify with a grep — so likely
-  output-stable in practice).
-- [x] **3.4 `CellMatcherLayerBitMaskTest`: out-of-bounds = no match.**
-  Today OOB probes clamp to the border cell, so `CellMatcherGridPattern`
-  patterns match at map edges against themselves. Make OOB return `false`.
-  `[output-changing]` at map borders — likely fixes visible edge artifacts;
-  verify against the sample generators.
-- [x] **3.5 `MarkerNodeActionPaintTerrain` mapping rebuild** on top of 3.1–3.2:
-  replace the `+ splat_to_grid` shift and hand-rolled region math with the
-  helpers; keep the Phase-1 clamping test green; add a "paint a radius-r marker
-  at center, splat 2× resolution → painted disc is centered and symmetric" spec.
-- [x] **3.6 `ContinuousGridCellActionSetTerrainHeight` coverage check.** With the
-  ratio convention, document that per-cell invocation (e.g. via
-  `ThemeEngine.applyCellRules`) writes only the texels under the cell footprint;
-  if the height sampler is ever denser than `2× grid`, callers must use
-  `GridActionRuleSet`'s `resolution` parameter. Add an assert comparing sampler
-  resolution to invocation resolution if cheap to thread through; otherwise doc.
-
----
-
-## Phase 4 — One initialization protocol
-
-Today: five signatures (`(grid, seed)`, `(grid, ecd, seed)`, `(seed, ecd, grid)`,
-`(data, ecd)`, `(grid)`), three flag disciplines, `ensureInitialized` used at 3 of
-~35 sites, `CellFilterCellMatcher` never sets its flag (re-initialized by every
-parent), and no composite finalizes recursively (stale noise tables/caches on
-reuse with a new grid/seed).
-
-Staged to keep the library API stable until the deliberate break:
-
-- [x] **4.1 (non-breaking) Base-class template within each hierarchy.**
-  For `CellFilter`: children enumerate via an overridable
-  `get children() { return EMPTY; }` on the base (Binary/Unary/Tertiary/complex
-  filters override with their inputs); `ensureInitialized` becomes the only
-  recursion driver (walks children, then calls the subclass hook); subclass
-  `initialize` overrides lose all flag/child bookkeeping; `finalize` recurses the
-  same way. Delete the ~15 hand-rolled `if (!x.initialized) x.initialize(...)`
-  blocks inside the filter hierarchy. Fixes the `CellFilterCellMatcher` flag bug
-  by construction. Same treatment inside `CellMatcher` (flagless today — give it
-  the same base machinery or document statelessness as a requirement).
-  Tests: shared-subtree DAG (one filter referenced by two parents) initialized
-  once per generation; re-initialization after `finalize` with a different seed
-  actually re-seeds (SimplexNoise spec: different seeds → different output after
-  finalize+initialize).
-- [x] **4.2 (breaking, one sweep) Signature unification.**
-  Decision first: target signature `initialize(grid, ecd, seed)` everywhere
-  (hierarchies that don't need `ecd` ignore it), or keep two sanctioned
-  signatures (`(grid, seed)` and `(grid, ecd, seed)`) and only fix the outliers
-  (`ContinuousGridCellAction`'s `(seed, ecd, grid)`, `Theme`/`CellProcessingRule*`'s
-  `(seed, ecd, grid)`, `MarkerNodeEntityProcessor`'s missing seed).
-  Recommendation: the two-signature option — smaller blast radius, still removes
-  the argument-order traps. Sweep all subclasses + meep samples + `app/src/generator`
-  call sites in one commit. Includes 2.8's `MarkerNodeEntityProcessor` seed.
-- [x] **4.3 `AbstractCellFilterSampleGridLayer`**: throw the same friendly
-  "Layer 'x' not found" error `GridLayerCellMatcher` throws instead of a
-  `TypeError` on the next line.
-
----
-
-## Phase 5 — Library architecture
-
-- [x] **5.1 Invert the `src → samples` dependency.** (Do early — 1.3's test wants it.)
-  `grid/generation/road/GridTaskGenerateRoads.js` imports `GridTags`,
-  `MirGridLayers`, `matcher_tag_traversable`; `grid/generation/discrete/GridTaskConnectRooms.js`
-  imports `matcher_tag_unoccupied`. A library must not hardcode one sample
-  game's tag scheme. Make `traversable`/`modifiable`/`actions` **required**
-  `from({...})` parameters (house rule: required params, not null-default
-  sentinels). Update the two classes, the samples, and `app/src/generator`
-  call sites. Roads also: `vertex` field is unused — remove or wire it; the
-  unused `weightMap` argument passed to `buildPaths` goes too.
-- [x] **5.2 One weighted-pick helper.** Extract the duplicated
-  cumulative-weights + `binarySearchHighIndex` selection from
-  `CellActionSelectWeightedRandom` and `MarkerNodeActionSelectWeighted` into a
-  typed core function (e.g. `random_pick_weighted_index(random, weights, count)`
-  over a caller-owned cumulative array). Both classes keep their typed public
-  shape; total-weight-0 handled once (see 1.12). Keep the two element-wrapper
-  classes — they're the typed API.
-- [ ] **5.3 One grid BFS/Dijkstra core.** Four hand-rolled variants:
-  `GridTaskBuildSourceDistanceMap`, `util/buildDistanceMapToObjective`,
-  `GridTaskConnectRooms.fillConnectedArea` (a BinaryHeap used as a FIFO),
-  roads' `tBuildPaths` loop. Extract one expansion routine with: multi-source
-  seeding, traversable predicate, per-visit callback / termination predicate,
-  and **mark-on-push** membership (BitSet) instead of `heap.contains` linear
-  scans (O(n²) today). Migrate the four call sites; behavior must be
-  output-stable (same expansion order — keep pop-order semantics identical;
-  verify with the roads/connect-rooms e2e tests from Phase 1).
-- [x] **5.4 `MarkerNode` position authority.** Nodes carry grid `position` and
-  world `transform`, hand-synced; `MarkerNodeTransformerOffsetPosition` moves
-  only grid, `…AddPositionYFromFilter` only world — after either, the other is
-  stale. Decision: grid `position` (+ yaw + scale) is authoritative; world XZ is
-  derived via `GridData.transform` at consumption time (entity placement,
-  PaintTerrain); world Y remains free (terrain cling / Y-offset transformers).
-  Concretely: `MarkerNodeActionEntityPlacement` recomputes world XZ from grid
-  position before composing transforms, and `OffsetPosition` stops being a
-  silent divergence point. Audit every `transform.position` reader in the
-  package + game call sites before flipping. This is the largest semantic
-  refactor in the plan — schedule last, behind a dedicated decision review.
-- [ ] **5.5 Task-ordering mechanisms.** `GridTaskGenerator.dependencies` resolves
-  by `indexOf` among direct siblings of one `GridTaskGroup` (cross-group deps
-  throw "not found"), while `GridTaskSequence` is structural and *rejects*
-  children with dependencies. Decision: (a) document the constraint and improve
-  the error message (cheap), or (b) resolve dependencies via a build-context map
-  (generator → task) threaded through nested `build` calls so cross-composite
-  deps work. Recommendation: (b) is a contained change to `GridTaskGroup`/
-  `GridTaskSequence.build` signatures (internal callers only) and removes the
-  landmine; do it when a real cross-group dependency first appears, (a) now.
-- [x] **5.6 Combinator hierarchies — explicit go/no-go.** Full unification
-  (matchers as boolean-typed filters; generic marker predicates) would remove
-  ~2 of the 4 And/Or/Not families and both bridge classes, but it is a breaking
-  conceptual change to the library's public vocabulary, and ruling #1 limits how
-  much boilerplate it can actually save. Default: **no-go for now.** Keep the
-  four families; rely on 1.1's full truth-table specs (replicated for each
-  family) to keep the hand-copied combinators honest. Revisit only alongside a
-  major-version API pass.
-- [x] **5.7 Non-functional stubs — finish or remove (decision, per ruling #2).**
-  - `grid/generation/discrete/layer/GridTaskDistanceToMarkers.js` — `build`
-    references an undefined `target` and returns `undefined`; crashes any
-    caller via `GridTaskGroup`. No caller can use it as-is.
-  - `grid/generation/grid/GridTaskGridAlignedNodeGenerator.js` — empty `build`
-    returning `undefined`; same failure mode.
-  - `GraphGenerationRule.js` / `GridGeneratorConfig.js` — inert data holders,
-    harmless; keep unless a vocabulary cleanup happens.
-  Options per stub: finish (BuildSourceDistanceMap covers DistanceToMarkers'
-  apparent intent — maybe it's already superseded), or remove. Owner call.
-- [x] **5.8 `GridTaskConnectRooms` / roads internal cleanups** (output-stable):
-  `fillConnectedArea` → stack/queue + mark-on-push (via 5.3); drop the
-  re-flood-of-everything per connected room if 5.3 makes incremental frontiers
-  cheap (measure first — only worth it on large grids); `drawPath`'s
-  constant `remap(0, n, 1, 1, i)` thickness scaffolding → either implement
-  per-endpoint thickness or simplify to the constant.
-
----
-
-## Testing & verification protocol
-
-- **Where/how:** Jest from `H:/git/moh` with the root `jest.conf.json`.
-  Black-box only: build a real `GridData` (+ real layers, markers, filters),
-  run the task/filter/action to completion, assert observable state (layer
-  contents, marker set, returned values). No mocks, no call-count assertions.
-  Test-local *recording* subclasses of action/consumer base classes are fine —
-  they observe behaviour through the public contract.
-- **Determinism suite (new, cheap, high value):** for each task generator fixed
-  in this plan, one spec asserting two runs with the same seed produce identical
-  grid/layer/marker state, and (where meaningful) different seeds differ.
-  This is the regression net for all `[output-changing]` work.
-- **`[output-changing]` verification:** after each Phase-2/3 batch, run
-  `SampleGenerator0` / `AlienGridGenerator` (and the combat generator) with a
-  fixed seed and eyeball the result against pre-change captures. Batch
-  output-changing merges so level-gen output shifts a known number of times,
-  not once per commit.
-- **Coverage debt:** the bugs found all lived in files with constructor-only or
-  no specs (`MarkerNodeMatcher*`, roads, `GridTaskExecuteRuleTimes`,
-  `ThemeEngine.applyNodes`, `PaintTerrain`, the CA). Every Phase-1 fix lands
-  with its spec; don't backfill beyond that until the phases above are done.
-
-## Explicitly rejected / out of scope (for the record)
-
-- String-keyed operation factories (`CellFilterBinary.of('add', …)`) — rejected;
-  typed verbosity preferred (ruling #1).
-- Deleting unreferenced-but-working classes (`CellFilterAnd`, `CellFilterSobel`,
-  `MarkerNodeEmitterGroup`, `NoopGridTaskGenerator`, etc.) — rejected; library
-  surface (ruling #2). Bugs inside them are fixed instead (1.9, 1.10, 1.11).
-- Any treatment of `prototype*.js` files — internal tooling, out of scope
-  (ruling #3).
-- Blanket replacement of module-static scratch pools — house style; assert
-  guards instead (1.14).
-
----
-
-## Execution log
-
-**Phase 1 + 5.1 (2026-06-11)** — all items landed with specs, full generation
-suite green (50 suites / 109 tests). Decisions taken during execution:
-
-- 1.4: exhaustion keeps signalling `EndFailure` (current intent preserved);
-  cells are now closed per attempt, so a rule can no longer fire twice on the
-  same cell within one build.
-- 1.12: the pre-existing spec pinned "single option with weight 0 still
-  executes" — flipped per plan to "all-zero weights execute nothing".
-- 1.14: guards added to `GridData.marker_collection` and `CellFilterCurvature`
-  scratch only; `CellFilterLookupTable.sample` and
-  `RoadConnectionNetwork.scratch_paths` run no user code between fill and read,
-  so they cannot re-enter — left unguarded.
-- 5.1: `GridTaskGenerateRoads` lost its unused `vertex` field and the dead
-  `weightMap` argument; `GridTaskConnectRooms.drawDebugState` now imports the
-  DOM helper lazily. Call sites updated: `SampleGenerator0` (meep),
-  `generateGameplay` (moh).
-- New test helper: `test_support/executeTaskTreeSync.js` — runs a Task/TaskGroup
-  tree synchronously, respecting dependencies.
-
-**Phase 2 (2026-06-11)** — semantic fixes landed, suite green (51 suites / 122
-tests). Decisions taken:
-
-- 2.1: OOB neighbours = dead (in this CA, alive = open space, so OOB-dead seals
-  borders); margins not frozen (same creep semantics as before); state is
-  double-buffered via front/back Uint8Arrays in GridTaskCellularAutomata.
-- 2.3: sigma = (samples-1)/6 in kernel-index units (kernel covers +-3 sigma).
-- 2.4: shared computeFilterSurfaceNormal helper; convention: z up, surface
-  rising along +x tilts normal towards -x; dome = positive curvature.
-- 2.5: one roll per (cell x rule x offset); rotation scan picks orientation
-  only. RNG is now drawn before matching, so draw sequences shift everywhere.
-- 2.6: aggregator normalize switched L2 -> L1 (terrain shader divides by weight
-  sum, so this is output-stable); zero-sum guard added; influence divisor
-  removed (proven no-op); resample uses texel-center mapping.
-- 2.7: estimation reads action.size directly; processArea builds nodes only
-  AFTER the overlap rejection (the spec caught this second consumption site);
-  testNodeCollisions uses new buildBaseNode (geometry only); purity
-  expectations documented on MarkerNodeTransformer.
-- 2.8: ArrayIteratorRandom seeding landed in 1.13; groupId documented as
-  opaque; derive_seed helper DEFERRED — applying it would shift all generation
-  output for marginal gain, revisit alongside a future intentional regen-shift;
-  EntityProcessor seed plumbed in Phase 4.
-
-**Phase 3 (2026-06-11)** — coordinate convention unified, suite green (52
-suites / 128 tests). Decisions taken:
-
-- Convention doc lives in `COORDINATES.md`; canonical helpers in `grid/coords/`.
-- DEVIATION from 3.3: `ContinuousGridCellActionSetTerrainHeight` was NOT
-  migrated — the terrain heightmap is the terrain renderer's vertex-aligned
-  domain and its corner mapping must match the terrain mesh, not grid layers.
-  Documented on the class and in COORDINATES.md (covers 3.6 as well).
-- 3.4: out-of-bounds layer probes now fail the match (border patterns stop
-  matching against themselves) — output-changing at map borders.
-- 3.5: PaintTerrain's one-texel center shift removed; painted discs are now
-  centered on the marker (verified at 1x and 2x splat resolution).
-
-**Phase 4 (2026-06-11)** — init protocol unified, suite green (53 suites / 130
-tests). Decisions taken:
-
-- 4.1: `CellFilter` base drives recursion via an overridable `get children()`;
-  `ensureInitialized` is the guarded entry, `initialize` the unconditional
-  subclass hook (super first), `finalize` recurses. All hand-rolled
-  `if (!x.initialized) x.initialize(...)` sites replaced with
-  `ensureInitialized`. `CellFilterCellMatcher`'s missing flag fixed by
-  construction. `CellMatcher` stays flagless: idempotence documented as the
-  contract.
-- 4.2: went with the two-sanctioned-signatures option — `(grid, seed)` and
-  `(grid, ecd, seed)`. Outliers flipped: `ContinuousGridCellAction` family
-  (+ `ensureInitialized` added to that base), `Theme`, `CellProcessingRule(Set)`,
-  `ThemeEngine.initializeThemes`. `MarkerNodeEntityProcessor.initialize` now
-  receives `seed`; `RandomRotation` finally varies with the world seed
-  (output-changing, was the point). `SetTerrainObstacle`'s threshold filter
-  is initialized via the same sweep.
-- New spec `filtering/CellFilter.spec.js` pins DAG init-once and
-  finalize-then-reseed.
-
-**Phase 5 (2026-06-11)** — suite green (53 suites / 130 tests). Status per item:
-
-- 5.2 DONE: `core/math/random/random_pick_weighted_index.js` — both weighted
-  selects rewired; zero-total guard lives in one place; RNG draw order
-  unchanged (helper does not consume the RNG when total weight is 0, same as
-  the phase-1 guards).
-- 5.3 PARTIAL (deliberate): mark-on-push membership BitSets replace every
-  linear `heap.contains` scan (BuildSourceDistanceMap, roads buildPaths), and
-  `fillConnectedArea` is a plain stack flood-fill instead of a BinaryHeap used
-  as a FIFO (order-independent result, verified by the e2e specs). The full
-  one-core-for-all extraction was NOT done: the four loops differ in exactly
-  the parts that matter (incremental task pumping, objective early-exit,
-  per-neighbor quadtree probes with early returns) — a shared core would be
-  callback soup costlier than the duplication. Revisit if a fifth variant
-  appears.
-- 5.4 RESOLVED (owner framing, 2026-06-11): `.position` is the virtual 2D-domain
-  location, `transform` is concrete world-space presentation — two domains, not
-  duplication; the "pick one authority" sweep is dead. Under that framing the
-  only defect was `MarkerNodeTransformerOffsetPosition` relocating the marker
-  in the virtual domain only; it now maintains the planar correspondence
-  (output-changing in the intended direction: offset wall segments / base
-  connectors finally move in world space — GenerateWalls, place_bases).
-  Domain contract documented on the `MarkerNode` fields.
-- 5.5 option (a) DONE: GridTaskGroup's dependency-resolution error now explains
-  the sibling-only constraint and points at GridTaskSequence. Option (b)
-  (build-context resolution map) deferred until a real cross-group dependency
-  appears.
-- 5.6 NO-GO confirmed: combinator hierarchies stay; truth-table specs (phase 1)
-  are the guard against future copy-paste drift.
-- 5.7 RESOLVED (owner: "if something is a stub - remove it"):
-  `GridTaskDistanceToMarkers` and `GridTaskGridAlignedNodeGenerator` removed.
-- 5.8 DONE: `drawPath` constant-thickness scaffolding simplified away (output
-  identical: thickness was constant 1); the per-room re-flood optimization
-  skipped per the item's "measure first".
-
-**Coverage pass (2026-06-11)** — package-wide first-principles black-box specs:
-117 suites / 234 tests (from 54 / 132). Constructor-only specs replaced with
-behavioural ones. Found and fixed in the process: `MarkerNode.copy` dropped
-`priority` — every cloning transformer silently reset a node's processing
-priority (applyNodes sorts on it).
-
-Deliberately NOT given dedicated specs: abstract bases and one-line data
-holders; debug tooling (`visualizeMarkers`, `visualise_filters_as_grid`,
-LogToConsole actions/filters, `GridCellActionDebugBreak` — contains a
-`debugger` statement); `MarkerNodeEntityProcessorClingToTerrain` and
-`ThemeEngine.applyTerrainThemes` (both need the built terrain tile/asset
-pipeline — integration tests, not headless unit territory);
-`GridTaskActionRuleSet` (one-line delegation, covered through
-GridActionRuleSet's own specs).
-
-VISUAL VERIFICATION PENDING: phases 1–4 change generated output (fixed seeds
-produce different levels). Run SampleGenerator0 / AlienGridGenerator and the
-combat generator against pre-change captures before shipping. Phase 5 and the
-coverage pass are output-stable (the priority fix only affects nodes that were
-losing priority through transformer clones).
+# Generation Package — Review 01 Action Plan
+
+Source: architectural + correctness review of `meep/src/generation`, 2026-06-11.
+Paths below are relative to this directory unless noted.
+
+## Scope rulings (from review discussion)
+
+These override the raw review findings:
+
+1. **No string-keyed factories.** One-class-per-operation stays. We accept the
+   verbosity of `CellFilterAdd` etc. in exchange for typed, greppable, JSDoc-able
+   code. Boilerplate reduction must come from better *base classes*, never from
+   `X.of('add', fn)`-style magic.
+2. **This is a library.** "Nothing imports it" is not an argument for removal.
+   Unreferenced classes are public surface: bugs in them get **fixed**, not
+   deleted. Removal is only on the table for code that *cannot work for any
+   caller* (non-functional stubs), and that is a per-item decision.
+3. **`prototype*.js` files are internal prototyping tools**, never published.
+   Out of scope for this plan.
+4. House rules that shape the fixes: correctness first (no bandaids), black-box
+   tests with real objects (no mocks, no call-count assertions), Jest runs from
+   the `moh` root with the root `jest.conf.json`, asserts are free in prod,
+   module-static pools are accepted style (guard them, don't remove them),
+   prefer required parameters over null-means-default sentinels.
+
+**Output-stability flag:** items marked `[output-changing]` alter generated
+levels for a fixed seed. Each such fix must be verified visually (see
+*Verification protocol* at the bottom) and batched consciously — they are the
+reason this plan is phased.
+
+---
+
+## Phase 1 — Surgical bug fixes
+
+Local, mechanically obvious fixes. Each lands with a black-box spec.
+
+- [x] **1.1 `MarkerNodeMatcherAnd.match` implements OR.**
+  `markers/matcher/MarkerNodeMatcherAnd.js`
+  Fix: `if (!this.left.match(node)) return false; return this.right.match(node);`
+  Test: spec with two `MarkerNodeMatcherByType`/`ContainsTag` matchers over a real
+  `MarkerNode`; truth-table all four combinations. Do the same table for `Or` and
+  `Not` while in there — the family has zero coverage.
+  `[output-changing]` — restores intended semantics in
+  `app/src/generator/generator/objectives/GENERATE_BOSS_OBJECTIVE.js` and
+  `generateMusic.js`. While fixing, read both call sites and confirm AND is what
+  they want (it is what they wrote).
+
+- [x] **1.2 Road relaxation typo.**
+  `grid/generation/road/GridTaskGenerateRoads.js:287`
+  Fix: `distances[neighbour_index] = distances;` → `= distance;`
+  Test: covered by the end-to-end roads test in 1.3 (the branch is hard to force
+  in isolation; the e2e test plus an `assert.isNumber` in the loop is enough).
+
+- [x] **1.3 Road groups self-connect and are then skipped.**
+  `grid/generation/road/GridTaskGenerateRoads.js` (node-encounter branch, ~line 336)
+  Fix: skip own-group nodes the same way the path-encounter branch skips own-group
+  paths: `if (targetGroupId === groupId) continue;` before the
+  `directGroupConnectionExists` check.
+  Test (e2e, black-box): small `GridData`, all-traversable layer, two marker
+  groups of **two adjacent connectors each**, placed apart; run the task to
+  completion; assert a `RoadConnection` exists whose endpoints belong to
+  *different* groups, and that every group is reachable in the connection graph.
+  Note: requires the matcher/action parameters from 5.1, or temporarily importing
+  the sample matchers in the spec (acceptable for the test, not for src).
+  Also in this file: delete the leftover `console.log(path)` in `tDrawPaths`, and
+  the constructed-but-unused `Graph` instance.
+
+- [x] **1.4 `GridTaskExecuteRuleTimes` never terminates when the pattern can't match.**
+  `grid/generation/GridTaskExecuteRuleTimes.js`
+  Fix: `closed.set(index, true)` after each attempted cell (match or not), and
+  make the exhaustion check count closed cells (`closedCount >= gridSize`)
+  instead of the linear-probe counter `i >= endIndex`.
+  Semantics note: this closes cells whose match could become true later as the
+  grid mutates. That is the original design intent (the `closed` set exists);
+  document it in the class doc.
+  Test: rule whose pattern never matches → task ends with failure within one
+  pass; rule that can match exactly `k < count` cells → terminates, `k`
+  placements. Used live in `AlienGridGenerator` / boss objective — confirm the
+  failure signal (`EndFailure`) is acceptable to those flows or whether
+  exhaustion should be `EndSuccess` with fewer placements. **Decision:** failure
+  vs. partial success on exhaustion.
+
+- [x] **1.6 `MarkerNodeTransformerYRotateByFilter` Euler convention mismatch.**
+  `markers/transform/MarkerNodeTransformerYRotateByFilter.js`
+  Fix: `toEulerAnglesXYZ` → `toEulerAnglesYXZ`, matching the recompose call and
+  the sibling `…YRotateByFilterGradient`.
+  Test: node with a known non-trivial prior YXZ rotation (x ≠ 0), constant
+  filter; assert resulting quaternion equals `fromEulerAnglesYXZ(x, expectedY, z)`.
+  `[output-changing]` for rotated decor in `theme-1/theme-2 injectDecor`.
+
+- [x] **1.7 `ThemeEngine.applyNodes` searches a stale-length influence array.**
+  `theme/ThemeEngine.js` (applyNodes)
+  Fix: pass explicit bounds to `binarySearchLowIndex(themeInfluence, t, cmp, 0, matchingThemeCount - 1)`
+  (mirroring the weighted selects). Handle `influenceSum === 0`: pick uniformly
+  among the matching themes using the same `random` stream (deterministic), since
+  "all themes at mask-distance 0" is a legitimate boundary state.
+  Test: grid with two overlapping `AreaTheme` masks; a marker matching 3 themes
+  processed immediately before one matching 1 theme; assert the second node is
+  processed by its only matching theme. Use a test-local recording
+  `MarkerNodeAction` subclass (a real action, observable behaviour — not a mock).
+  `[output-changing]` at theme boundaries.
+
+- [x] **1.8 `MarkerNodeActionPaintTerrain`: clamp the splat region.**
+  `markers/actions/terrain/MarkerNodeActionPaintTerrain.js`
+  Fix (this phase, pure bug): clamp `x0/x1/y0/y1` to `[0, splat_resolution-1]`
+  so edge markers stop wrapping writes into the opposite row. Make the X/Y
+  region rounding symmetric (`floor(v + 0.5)` both axes).
+  The deeper grid↔texel mapping question (the `+ splat_to_grid` "cell center"
+  shift) is deferred to Phase 3 — leave a `TODO(review-01 §3)` comment.
+  Test: marker at grid (0,0) with radius reaching off-map; assert no texel in the
+  rightmost column / other rows changes. `[output-changing]` within one texel of
+  map edges only.
+
+- [x] **1.9 `ContinuousGridCellActionWriteObstacle` writes global coords.**
+  `grid/actions/ContinuousGridCellActionWriteObstacle.js`
+  Fix: `to.writePoint(local_x, local_y, …)`. Library surface — fix despite no
+  internal callers. Test: obstacle target with non-zero `GridPosition`; execute at
+  a global coordinate; assert the *local* cell is written.
+
+- [x] **1.10 `MarkerNode.overlaps` compares the wrong plane.**
+  `markers/MarkerNode.js`
+  Fix: use `this.position` / `other.position` (grid-space Vector2), matching
+  `GridDataNodePredicateOverlaps`. Library surface; fix, add a 3-case spec
+  (overlap / touch / disjoint).
+
+- [x] **1.11 `CellFilterLiteralBoolean` cannot execute.**
+  `filtering/boolean/CellFilterLiteralBoolean.js`
+  Fix: implement `execute() { return this.value ? 1 : 0; }` (the package-wide
+  boolean-as-number convention, same as `CellFilterCellMatcher`). Add `TRUE` /
+  `FALSE` singletons alongside `CellFilterLiteralFloat.ONE/ZERO` for symmetry.
+  Spec: both values, plus `dataType === DataType.Boolean`.
+
+- [x] **1.12 Weighted selects: total weight 0 still executes an element.**
+  `placement/action/random/weighted/CellActionSelectWeightedRandom.js`,
+  `markers/actions/probability/MarkerNodeActionSelectWeighted.js`
+  Fix: if `totalWeight <= 0`, execute nothing. (Folded into 5.2's shared helper;
+  if 5.2 is deferred, fix both copies now.)
+  Test: all-zero weight filters → no action executed (recording action).
+
+- [x] **1.13 Small fixes batch** (no individual specs needed beyond existing suites):
+  - `GridTaskGroup.catchGeneratorErrors` — include `reason` in the
+    `console.error` (currently drops the actual error).
+  - `GridTaskApplyActionToCells.from` — apply the `name` parameter; rename the
+    misleading `fitness` field to `supplier` (or `selector`) to match the API.
+  - `CellSupplierBestN` — implement `count()` (`this.elements.length`) so task
+    progress works.
+  - `GridTaskGenerateRoads` — assert integer connector positions before
+    `index = x + y * width` (fractional positions silently corrupt the search).
+  - `GridData.resize` — assert message typo (`'width'` on the height check).
+  - `CellFilterInverseLerp.from` — JSDoc for `f`/`value` is copy-pasted from Lerp.
+  - `sampler_from_filter` — assert says non-negative, doc says positive; make it
+    `greaterThan(0)`.
+  - `GridTaskConnectRooms` — move the unconditional `drawSamplerHTML` import into
+    the (commented) debug path or a lazy import so headless bundles don't pull
+    DOM helpers. Keep `debugState`/`drawDebugState` (internal tooling).
+  - `GridActionRuleSet.process` — seed the `Random` policy iterator
+    (`ArrayIteratorRandom.setSeed(seed)`) so rule order responds to the world
+    seed. `[output-changing]` for `Random`-policy rule sets.
+
+- [x] **1.14 Re-entrancy guards on module-static pools.**
+  Module-static pools are house style and stay. Add cheap in-use assert guards
+  (free in prod) where recursion through user code is possible:
+  `GridData.marker_collection` (matcher could re-query the grid),
+  `CellFilterCurvature`'s normal scratch (curvature-of-curvature),
+  `CellFilterLookupTable.sample`, `RoadConnectionNetwork.scratch_paths`.
+  Pattern: `assert.ok(!pool_in_use, 're-entrant use of <pool>')` + try/finally
+  flag toggle is NOT acceptable (silent catch rule) — use a plain set/unset; an
+  assert firing in dev is the goal.
+
+---
+
+## Phase 2 — Semantic decisions, then fixes
+
+Each item changes generated output by design. Decide the intended semantics,
+fix, then verify via the sample generators (see Verification protocol).
+
+- [x] **2.1 Cellular automata: synchronous update.**
+  `automata/CaveGeneratorCellularAutomata.js`, `grid/generation/discrete/GridTaskCellularAutomata.js`
+  Current `step` mutates in place → sequential CA with scan-order bias.
+  Decision A: change `step(data, …)` to `step(src, dst, …)` and ping-pong the two
+  buffers across the 50 iterations in `GridTaskCellularAutomata` (signature
+  change to the `CellularAutomata` base — library surface, sweep subclasses).
+  Decision B (same time): out-of-bounds neighbor policy. Current clamping
+  double-counts edge cells; the conventional cave-CA choice is OOB = wall
+  (alive), which seals cave borders. Margin seeding interaction: margins start
+  dead but can currently come alive — decide whether margins should be frozen.
+  Also: write `0/1`, not `false/true`, into the `Uint8Array`.
+  Test: synchronicity — a glider-style asymmetric seed must produce the same
+  result as a reference two-buffer implementation; plus determinism (same seed →
+  identical field). `[output-changing]` — all CA-generated caves.
+
+- [x] **2.2 Fractal noise octave direction.**
+  `filtering/numeric/complex/CellFilterSimplexNoise.js` (`fractal`)
+  `scale` is a wavelength; passing `scale * frequency` makes successive octaves
+  *lower* frequency at *lower* amplitude — inverted fBm. Intended:
+  `scale / frequency`.
+  Decision: fix to conventional fBm, or keep current output and rename/document?
+  Recommendation: fix; the parameter names (`octaves`, `persistence`,
+  `lacunarity`) promise conventional fBm. Grep `fractal(` call sites first and
+  eyeball each generated surface after.
+  Test: spectral sanity — with 3 octaves, correlation length of the output must
+  *decrease* as octave count rises (or simpler: octave-2 filter alone must have
+  shorter wavelength than octave-1; assert via zero-crossing counts along a row).
+  `[output-changing]`.
+
+- [x] **2.3 Gaussian blur sigma units.**
+  `filtering/numeric/complex/CellFilterGaussianBlur.js`
+  `sigma = radius * 3` in kernel-index units makes the kernel flat (a box blur).
+  For a real Gaussian: sigma ≈ kernel half-width / 3, i.e.
+  `sigma = (samples - 1) / 6` in index units.
+  Decision: is the box-like character relied upon visually? If yes, keep and
+  rename the comment; if no, fix the constant. Recommendation: fix, since the
+  class advertises Gaussian and callers tuned radius, not shape.
+  Test: kernel center weight vs. edge weight ratio > some threshold (e.g. center
+  ≥ 3× edge for quality 3) via `build_gaussian_kernel_2d` inspection — that's
+  testing the helper's contract, not implementation internals.
+  `[output-changing]` — all blurred filters.
+
+- [x] **2.4 Heightfield normal convention (sign of gradient).**
+  `filtering/numeric/complex/CellFilterAngleToNormal.js`,
+  `filtering/numeric/complex/CellFilterCurvature.js`
+  Both build normals as `(+dX, +dY, 2)`; a true surface normal is
+  `∝ (−dX, −dY, 2)`. AngleToNormal is unaffected for the default vertical
+  reference; directional references get mirrored slopes; Curvature's sign is
+  inverted vs. its cited reference shader.
+  Fix: extract one shared `computeFilterSurfaceNormal(result, filter, grid, x, y)`
+  helper (also de-duplicates the two implementations), with the negated-gradient
+  convention, and document the axis mapping (grid x,y,height → normal x,y,z).
+  Audit the two consumers for baked-in compensations before flipping.
+  Test: synthetic ramp filter (`h = x`): normal must tilt *toward −x*; curvature
+  of a paraboloid bump must be positive at the peak (pick and document the
+  convention).
+  `[output-changing]` where non-default references / curvature thresholds are used.
+
+- [x] **2.5 `GridActionRuleSet` rotation/probability coupling.**
+  `markers/GridActionRuleSet.js`
+  Today each rotation re-rolls probability: symmetric patterns get effective
+  `1−(1−p)⁴` and rotation 0 is favored.
+  Decision: intended semantics. Recommendation: roll once per (cell × rule ×
+  pattern-offset), then scan rotations for the first match — "does this rule
+  fire here?" and "in which orientation?" become independent.
+  Test: probability 0.5 rule, rotation-symmetric matcher, large grid, fixed
+  seed: firing rate within tolerance of 0.5 (black-box statistical bound);
+  rotation histogram roughly uniform across matching rotations if that's the
+  chosen semantic. `[output-changing]`.
+
+- [x] **2.6 ThemeEngine blending math.**
+  `theme/ThemeEngine.js`, `theme/TerrainLayerRuleAggregator.js`
+  - Remove the `influence = distance / matchingThemeCount` divisor — it cancels
+    in both consumers (verified no-op; pure simplification, output-stable).
+  - `aggregator.normalize(255)` normalizes by **L2**; splat weights conventionally
+    sum to 255 (L1). Investigate what the terrain shader expects before changing.
+    If shader renormalizes, L2→L1 is output-stable; otherwise `[output-changing]`.
+  - Guard `d === 0` (no theme contributes) explicitly instead of relying on
+    `NaN → uint8 0`.
+  - `tResample` half-texel: source coords should be
+    `(x + 0.5) / splatWidth * weights.width − 0.5`; decide if the visual shift
+    is worth fixing while in here. `[output-changing]` (sub-texel).
+
+- [x] **2.7 Stateful transformers vs. estimation/test paths.**
+  `grid/generation/GridTaskDensityMarkerDistribution.js` (`estimateTapCount`),
+  `markers/GridCellActionPlaceMarkerGroup.js` (`testNodeCollisions`),
+  `markers/transform/MarkerNodeTransformerRecordUniqueRandomEnum.js`
+  `buildNode` runs transformers; estimation/collision-test paths therefore
+  *consume* from stateful transformers (UniqueRandomEnum's pool drains before
+  real placement).
+  Fixes, in order of preference:
+  1. `estimateTapCount` doesn't need nodes at all — it uses only `node.size`.
+     Compute size as `action.size * scale.sampleRandom(random)` directly; no
+     `buildNode` call. (Current transformers never alter `size`; assert that
+     invariant in `buildNode` if we rely on it — or just accept estimate drift.)
+  2. `testNodeCollisions`: grep callers; it wants post-offset positions, so it
+     can't skip transformers blindly. Split `buildNode` into the geometric phase
+     (offset/rotation/transform composition) and the transformer phase as two
+     named methods; `testNodeCollisions` uses the geometric phase. Document the
+     residual inaccuracy for position-moving transformers, or accept running
+     transformers there and document the purity requirement.
+  3. Document on `MarkerNodeTransformer`: transformers should be pure w.r.t.
+     their own state per `transform` call *or* tolerate being called for
+     discarded nodes. `RecordUniqueRandomEnum` inherently can't — note it in its
+     class doc as incompatible with estimating tasks.
+  Test: DensityMarkerDistribution with a UniqueRandomEnum transformer whose pool
+  is exactly the expected placement count → must not throw "All values used up".
+
+- [x] **2.8 Seed plumbing decisions.**
+  - `MarkerNodeEntityProcessorRandomRotation` ignores the world seed because
+    `MarkerNodeEntityProcessor.initialize(data, ecd)` has no seed parameter.
+    Add `seed` to the processor interface (sweep: ClingToTerrain, Sequence,
+    RandomRotation). Folds into Phase 4's signature unification — do it there.
+    `[output-changing]` (entity rotations start varying with seed — that's the point).
+  - `GridCellActionPlaceMarkerGroup`'s module-global `idCounter`: uniqueness is
+    all that's required today, but it breaks cross-session reproducibility of
+    `properties.groupId` values. Decision: leave (document "IDs are opaque,
+    equality-only") or derive from a per-generation counter owned by the task.
+    Recommendation: document as opaque now; revisit if save/replay needs arise.
+  - Standardize derived seeds (`seed + i`, `(s+1)*31`, `seed + 91234`, …) on one
+    small typed helper (e.g. `derive_seed(seed, salt)`) — mechanical, low risk,
+    `[output-changing]` wherever applied, so batch it with other output-changing
+    work, not piecemeal.
+
+---
+
+## Phase 3 — One grid↔sampler coordinate convention
+
+The package has three conversions for the same mapping: corner-aligned
+`x/(w−1)·(sw−1)` (`GridCellActionWriteFilterToLayer`, `CellMatcherLayerBitMaskTest`,
+`ContinuousGridCellActionSetTerrainHeight`, `populateSampler2DFromCellFilter`),
+ratio `x·(sw/w)` (`AbstractCellFilterSampleGridLayer` → both sample-layer
+filters), and PaintTerrain's bespoke shift. Writes and reads through different
+conventions drift by up to a cell at the far edge whenever `resolution ≠ 1`,
+and `/(w−1)` divides by zero on 1-wide grids.
+
+- [x] **3.1 Write the mapping spec.** One short doc section (in this file or a
+  `COORDINATES.md`): grid integer coordinates are cell centers; layer texel `t`
+  covers grid `[t/r, (t+1)/r)`; conversion is the ratio form
+  `texel = grid · resolution` (consistent with `GridDataLayer.resolution` and
+  `GridData.computeScale`'s half-tile world offset). Define the out-of-bounds
+  policy per operation class: *reads* clamp or return a defined exterior value
+  (pick one), *matches* fail (no silent edge-clamping), *writes* are discarded
+  with bounds checks.
+- [x] **3.2 Implement typed helpers** (plain functions, no string keys):
+  `grid_to_texel(x, resolution)`, `texel_to_grid(t, resolution)` or equivalent —
+  small enough to inline, but one canonical definition to reference.
+- [x] **3.3 Migrate the corner-convention sites** listed above to the ratio
+  convention. Each migration gets a `resolution = 2` round-trip spec
+  (write filter → layer → read via `CellFilterSampleLayerLinear`, assert values
+  at interior and far-edge cells). `[output-changing]` for `resolution ≠ 1`
+  layers (game currently uses resolution 1 — verify with a grep — so likely
+  output-stable in practice).
+- [x] **3.4 `CellMatcherLayerBitMaskTest`: out-of-bounds = no match.**
+  Today OOB probes clamp to the border cell, so `CellMatcherGridPattern`
+  patterns match at map edges against themselves. Make OOB return `false`.
+  `[output-changing]` at map borders — likely fixes visible edge artifacts;
+  verify against the sample generators.
+- [x] **3.5 `MarkerNodeActionPaintTerrain` mapping rebuild** on top of 3.1–3.2:
+  replace the `+ splat_to_grid` shift and hand-rolled region math with the
+  helpers; keep the Phase-1 clamping test green; add a "paint a radius-r marker
+  at center, splat 2× resolution → painted disc is centered and symmetric" spec.
+- [x] **3.6 `ContinuousGridCellActionSetTerrainHeight` coverage check.** With the
+  ratio convention, document that per-cell invocation (e.g. via
+  `ThemeEngine.applyCellRules`) writes only the texels under the cell footprint;
+  if the height sampler is ever denser than `2× grid`, callers must use
+  `GridActionRuleSet`'s `resolution` parameter. Add an assert comparing sampler
+  resolution to invocation resolution if cheap to thread through; otherwise doc.
+
+---
+
+## Phase 4 — One initialization protocol
+
+Today: five signatures (`(grid, seed)`, `(grid, ecd, seed)`, `(seed, ecd, grid)`,
+`(data, ecd)`, `(grid)`), three flag disciplines, `ensureInitialized` used at 3 of
+~35 sites, `CellFilterCellMatcher` never sets its flag (re-initialized by every
+parent), and no composite finalizes recursively (stale noise tables/caches on
+reuse with a new grid/seed).
+
+Staged to keep the library API stable until the deliberate break:
+
+- [x] **4.1 (non-breaking) Base-class template within each hierarchy.**
+  For `CellFilter`: children enumerate via an overridable
+  `get children() { return EMPTY; }` on the base (Binary/Unary/Tertiary/complex
+  filters override with their inputs); `ensureInitialized` becomes the only
+  recursion driver (walks children, then calls the subclass hook); subclass
+  `initialize` overrides lose all flag/child bookkeeping; `finalize` recurses the
+  same way. Delete the ~15 hand-rolled `if (!x.initialized) x.initialize(...)`
+  blocks inside the filter hierarchy. Fixes the `CellFilterCellMatcher` flag bug
+  by construction. Same treatment inside `CellMatcher` (flagless today — give it
+  the same base machinery or document statelessness as a requirement).
+  Tests: shared-subtree DAG (one filter referenced by two parents) initialized
+  once per generation; re-initialization after `finalize` with a different seed
+  actually re-seeds (SimplexNoise spec: different seeds → different output after
+  finalize+initialize).
+- [x] **4.2 (breaking, one sweep) Signature unification.**
+  Decision first: target signature `initialize(grid, ecd, seed)` everywhere
+  (hierarchies that don't need `ecd` ignore it), or keep two sanctioned
+  signatures (`(grid, seed)` and `(grid, ecd, seed)`) and only fix the outliers
+  (`ContinuousGridCellAction`'s `(seed, ecd, grid)`, `Theme`/`CellProcessingRule*`'s
+  `(seed, ecd, grid)`, `MarkerNodeEntityProcessor`'s missing seed).
+  Recommendation: the two-signature option — smaller blast radius, still removes
+  the argument-order traps. Sweep all subclasses + meep samples + `app/src/generator`
+  call sites in one commit. Includes 2.8's `MarkerNodeEntityProcessor` seed.
+- [x] **4.3 `AbstractCellFilterSampleGridLayer`**: throw the same friendly
+  "Layer 'x' not found" error `GridLayerCellMatcher` throws instead of a
+  `TypeError` on the next line.
+
+---
+
+## Phase 5 — Library architecture
+
+- [x] **5.1 Invert the `src → samples` dependency.** (Do early — 1.3's test wants it.)
+  `grid/generation/road/GridTaskGenerateRoads.js` imports `GridTags`,
+  `MirGridLayers`, `matcher_tag_traversable`; `grid/generation/discrete/GridTaskConnectRooms.js`
+  imports `matcher_tag_unoccupied`. A library must not hardcode one sample
+  game's tag scheme. Make `traversable`/`modifiable`/`actions` **required**
+  `from({...})` parameters (house rule: required params, not null-default
+  sentinels). Update the two classes, the samples, and `app/src/generator`
+  call sites. Roads also: `vertex` field is unused — remove or wire it; the
+  unused `weightMap` argument passed to `buildPaths` goes too.
+- [x] **5.2 One weighted-pick helper.** Extract the duplicated
+  cumulative-weights + `binarySearchHighIndex` selection from
+  `CellActionSelectWeightedRandom` and `MarkerNodeActionSelectWeighted` into a
+  typed core function (e.g. `random_pick_weighted_index(random, weights, count)`
+  over a caller-owned cumulative array). Both classes keep their typed public
+  shape; total-weight-0 handled once (see 1.12). Keep the two element-wrapper
+  classes — they're the typed API.
+- [ ] **5.3 One grid BFS/Dijkstra core.** Four hand-rolled variants:
+  `GridTaskBuildSourceDistanceMap`, `util/buildDistanceMapToObjective`,
+  `GridTaskConnectRooms.fillConnectedArea` (a BinaryHeap used as a FIFO),
+  roads' `tBuildPaths` loop. Extract one expansion routine with: multi-source
+  seeding, traversable predicate, per-visit callback / termination predicate,
+  and **mark-on-push** membership (BitSet) instead of `heap.contains` linear
+  scans (O(n²) today). Migrate the four call sites; behavior must be
+  output-stable (same expansion order — keep pop-order semantics identical;
+  verify with the roads/connect-rooms e2e tests from Phase 1).
+- [x] **5.4 `MarkerNode` position authority.** Nodes carry grid `position` and
+  world `transform`, hand-synced; `MarkerNodeTransformerOffsetPosition` moves
+  only grid, `…AddPositionYFromFilter` only world — after either, the other is
+  stale. Decision: grid `position` (+ yaw + scale) is authoritative; world XZ is
+  derived via `GridData.transform` at consumption time (entity placement,
+  PaintTerrain); world Y remains free (terrain cling / Y-offset transformers).
+  Concretely: `MarkerNodeActionEntityPlacement` recomputes world XZ from grid
+  position before composing transforms, and `OffsetPosition` stops being a
+  silent divergence point. Audit every `transform.position` reader in the
+  package + game call sites before flipping. This is the largest semantic
+  refactor in the plan — schedule last, behind a dedicated decision review.
+- [ ] **5.5 Task-ordering mechanisms.** `GridTaskGenerator.dependencies` resolves
+  by `indexOf` among direct siblings of one `GridTaskGroup` (cross-group deps
+  throw "not found"), while `GridTaskSequence` is structural and *rejects*
+  children with dependencies. Decision: (a) document the constraint and improve
+  the error message (cheap), or (b) resolve dependencies via a build-context map
+  (generator → task) threaded through nested `build` calls so cross-composite
+  deps work. Recommendation: (b) is a contained change to `GridTaskGroup`/
+  `GridTaskSequence.build` signatures (internal callers only) and removes the
+  landmine; do it when a real cross-group dependency first appears, (a) now.
+- [x] **5.6 Combinator hierarchies — explicit go/no-go.** Full unification
+  (matchers as boolean-typed filters; generic marker predicates) would remove
+  ~2 of the 4 And/Or/Not families and both bridge classes, but it is a breaking
+  conceptual change to the library's public vocabulary, and ruling #1 limits how
+  much boilerplate it can actually save. Default: **no-go for now.** Keep the
+  four families; rely on 1.1's full truth-table specs (replicated for each
+  family) to keep the hand-copied combinators honest. Revisit only alongside a
+  major-version API pass.
+- [x] **5.7 Non-functional stubs — finish or remove (decision, per ruling #2).**
+  - `grid/generation/discrete/layer/GridTaskDistanceToMarkers.js` — `build`
+    references an undefined `target` and returns `undefined`; crashes any
+    caller via `GridTaskGroup`. No caller can use it as-is.
+  - `grid/generation/grid/GridTaskGridAlignedNodeGenerator.js` — empty `build`
+    returning `undefined`; same failure mode.
+  - `GraphGenerationRule.js` / `GridGeneratorConfig.js` — inert data holders,
+    harmless; keep unless a vocabulary cleanup happens.
+  Options per stub: finish (BuildSourceDistanceMap covers DistanceToMarkers'
+  apparent intent — maybe it's already superseded), or remove. Owner call.
+- [x] **5.8 `GridTaskConnectRooms` / roads internal cleanups** (output-stable):
+  `fillConnectedArea` → stack/queue + mark-on-push (via 5.3); drop the
+  re-flood-of-everything per connected room if 5.3 makes incremental frontiers
+  cheap (measure first — only worth it on large grids); `drawPath`'s
+  constant `remap(0, n, 1, 1, i)` thickness scaffolding → either implement
+  per-endpoint thickness or simplify to the constant.
+
+---
+
+## Testing & verification protocol
+
+- **Where/how:** Jest from `H:/git/moh` with the root `jest.conf.json`.
+  Black-box only: build a real `GridData` (+ real layers, markers, filters),
+  run the task/filter/action to completion, assert observable state (layer
+  contents, marker set, returned values). No mocks, no call-count assertions.
+  Test-local *recording* subclasses of action/consumer base classes are fine —
+  they observe behaviour through the public contract.
+- **Determinism suite (new, cheap, high value):** for each task generator fixed
+  in this plan, one spec asserting two runs with the same seed produce identical
+  grid/layer/marker state, and (where meaningful) different seeds differ.
+  This is the regression net for all `[output-changing]` work.
+- **`[output-changing]` verification:** after each Phase-2/3 batch, run
+  `SampleGenerator0` / `AlienGridGenerator` (and the combat generator) with a
+  fixed seed and eyeball the result against pre-change captures. Batch
+  output-changing merges so level-gen output shifts a known number of times,
+  not once per commit.
+- **Coverage debt:** the bugs found all lived in files with constructor-only or
+  no specs (`MarkerNodeMatcher*`, roads, `GridTaskExecuteRuleTimes`,
+  `ThemeEngine.applyNodes`, `PaintTerrain`, the CA). Every Phase-1 fix lands
+  with its spec; don't backfill beyond that until the phases above are done.
+
+## Explicitly rejected / out of scope (for the record)
+
+- String-keyed operation factories (`CellFilterBinary.of('add', …)`) — rejected;
+  typed verbosity preferred (ruling #1).
+- Deleting unreferenced-but-working classes (`CellFilterAnd`, `CellFilterSobel`,
+  `MarkerNodeEmitterGroup`, `NoopGridTaskGenerator`, etc.) — rejected; library
+  surface (ruling #2). Bugs inside them are fixed instead (1.9, 1.10, 1.11).
+- Any treatment of `prototype*.js` files — internal tooling, out of scope
+  (ruling #3).
+- Blanket replacement of module-static scratch pools — house style; assert
+  guards instead (1.14).
+
+---
+
+## Execution log
+
+**Phase 1 + 5.1 (2026-06-11)** — all items landed with specs, full generation
+suite green (50 suites / 109 tests). Decisions taken during execution:
+
+- 1.4: exhaustion keeps signalling `EndFailure` (current intent preserved);
+  cells are now closed per attempt, so a rule can no longer fire twice on the
+  same cell within one build.
+- 1.12: the pre-existing spec pinned "single option with weight 0 still
+  executes" — flipped per plan to "all-zero weights execute nothing".
+- 1.14: guards added to `GridData.marker_collection` and `CellFilterCurvature`
+  scratch only; `CellFilterLookupTable.sample` and
+  `RoadConnectionNetwork.scratch_paths` run no user code between fill and read,
+  so they cannot re-enter — left unguarded.
+- 5.1: `GridTaskGenerateRoads` lost its unused `vertex` field and the dead
+  `weightMap` argument; `GridTaskConnectRooms.drawDebugState` now imports the
+  DOM helper lazily. Call sites updated: `SampleGenerator0` (meep),
+  `generateGameplay` (moh).
+- New test helper: `test_support/executeTaskTreeSync.js` — runs a Task/TaskGroup
+  tree synchronously, respecting dependencies.
+
+**Phase 2 (2026-06-11)** — semantic fixes landed, suite green (51 suites / 122
+tests). Decisions taken:
+
+- 2.1: OOB neighbours = dead (in this CA, alive = open space, so OOB-dead seals
+  borders); margins not frozen (same creep semantics as before); state is
+  double-buffered via front/back Uint8Arrays in GridTaskCellularAutomata.
+- 2.3: sigma = (samples-1)/6 in kernel-index units (kernel covers +-3 sigma).
+- 2.4: shared computeFilterSurfaceNormal helper; convention: z up, surface
+  rising along +x tilts normal towards -x; dome = positive curvature.
+- 2.5: one roll per (cell x rule x offset); rotation scan picks orientation
+  only. RNG is now drawn before matching, so draw sequences shift everywhere.
+- 2.6: aggregator normalize switched L2 -> L1 (terrain shader divides by weight
+  sum, so this is output-stable); zero-sum guard added; influence divisor
+  removed (proven no-op); resample uses texel-center mapping.
+- 2.7: estimation reads action.size directly; processArea builds nodes only
+  AFTER the overlap rejection (the spec caught this second consumption site);
+  testNodeCollisions uses new buildBaseNode (geometry only); purity
+  expectations documented on MarkerNodeTransformer.
+- 2.8: ArrayIteratorRandom seeding landed in 1.13; groupId documented as
+  opaque; derive_seed helper DEFERRED — applying it would shift all generation
+  output for marginal gain, revisit alongside a future intentional regen-shift;
+  EntityProcessor seed plumbed in Phase 4.
+
+**Phase 3 (2026-06-11)** — coordinate convention unified, suite green (52
+suites / 128 tests). Decisions taken:
+
+- Convention doc lives in `COORDINATES.md`; canonical helpers in `grid/coords/`.
+- DEVIATION from 3.3: `ContinuousGridCellActionSetTerrainHeight` was NOT
+  migrated — the terrain heightmap is the terrain renderer's vertex-aligned
+  domain and its corner mapping must match the terrain mesh, not grid layers.
+  Documented on the class and in COORDINATES.md (covers 3.6 as well).
+- 3.4: out-of-bounds layer probes now fail the match (border patterns stop
+  matching against themselves) — output-changing at map borders.
+- 3.5: PaintTerrain's one-texel center shift removed; painted discs are now
+  centered on the marker (verified at 1x and 2x splat resolution).
+
+**Phase 4 (2026-06-11)** — init protocol unified, suite green (53 suites / 130
+tests). Decisions taken:
+
+- 4.1: `CellFilter` base drives recursion via an overridable `get children()`;
+  `ensureInitialized` is the guarded entry, `initialize` the unconditional
+  subclass hook (super first), `finalize` recurses. All hand-rolled
+  `if (!x.initialized) x.initialize(...)` sites replaced with
+  `ensureInitialized`. `CellFilterCellMatcher`'s missing flag fixed by
+  construction. `CellMatcher` stays flagless: idempotence documented as the
+  contract.
+- 4.2: went with the two-sanctioned-signatures option — `(grid, seed)` and
+  `(grid, ecd, seed)`. Outliers flipped: `ContinuousGridCellAction` family
+  (+ `ensureInitialized` added to that base), `Theme`, `CellProcessingRule(Set)`,
+  `ThemeEngine.initializeThemes`. `MarkerNodeEntityProcessor.initialize` now
+  receives `seed`; `RandomRotation` finally varies with the world seed
+  (output-changing, was the point). `SetTerrainObstacle`'s threshold filter
+  is initialized via the same sweep.
+- New spec `filtering/CellFilter.spec.js` pins DAG init-once and
+  finalize-then-reseed.
+
+**Phase 5 (2026-06-11)** — suite green (53 suites / 130 tests). Status per item:
+
+- 5.2 DONE: `core/math/random/random_pick_weighted_index.js` — both weighted
+  selects rewired; zero-total guard lives in one place; RNG draw order
+  unchanged (helper does not consume the RNG when total weight is 0, same as
+  the phase-1 guards).
+- 5.3 PARTIAL (deliberate): mark-on-push membership BitSets replace every
+  linear `heap.contains` scan (BuildSourceDistanceMap, roads buildPaths), and
+  `fillConnectedArea` is a plain stack flood-fill instead of a BinaryHeap used
+  as a FIFO (order-independent result, verified by the e2e specs). The full
+  one-core-for-all extraction was NOT done: the four loops differ in exactly
+  the parts that matter (incremental task pumping, objective early-exit,
+  per-neighbor quadtree probes with early returns) — a shared core would be
+  callback soup costlier than the duplication. Revisit if a fifth variant
+  appears.
+- 5.4 RESOLVED (owner framing, 2026-06-11): `.position` is the virtual 2D-domain
+  location, `transform` is concrete world-space presentation — two domains, not
+  duplication; the "pick one authority" sweep is dead. Under that framing the
+  only defect was `MarkerNodeTransformerOffsetPosition` relocating the marker
+  in the virtual domain only; it now maintains the planar correspondence
+  (output-changing in the intended direction: offset wall segments / base
+  connectors finally move in world space — GenerateWalls, place_bases).
+  Domain contract documented on the `MarkerNode` fields.
+- 5.5 option (a) DONE: GridTaskGroup's dependency-resolution error now explains
+  the sibling-only constraint and points at GridTaskSequence. Option (b)
+  (build-context resolution map) deferred until a real cross-group dependency
+  appears.
+- 5.6 NO-GO confirmed: combinator hierarchies stay; truth-table specs (phase 1)
+  are the guard against future copy-paste drift.
+- 5.7 RESOLVED (owner: "if something is a stub - remove it"):
+  `GridTaskDistanceToMarkers` and `GridTaskGridAlignedNodeGenerator` removed.
+- 5.8 DONE: `drawPath` constant-thickness scaffolding simplified away (output
+  identical: thickness was constant 1); the per-room re-flood optimization
+  skipped per the item's "measure first".
+
+**Coverage pass (2026-06-11)** — package-wide first-principles black-box specs:
+117 suites / 234 tests (from 54 / 132). Constructor-only specs replaced with
+behavioural ones. Found and fixed in the process: `MarkerNode.copy` dropped
+`priority` — every cloning transformer silently reset a node's processing
+priority (applyNodes sorts on it).
+
+Deliberately NOT given dedicated specs: abstract bases and one-line data
+holders; debug tooling (`visualizeMarkers`, `visualise_filters_as_grid`,
+LogToConsole actions/filters, `GridCellActionDebugBreak` — contains a
+`debugger` statement); `MarkerNodeEntityProcessorClingToTerrain` and
+`ThemeEngine.applyTerrainThemes` (both need the built terrain tile/asset
+pipeline — integration tests, not headless unit territory);
+`GridTaskActionRuleSet` (one-line delegation, covered through
+GridActionRuleSet's own specs).
+
+VISUAL VERIFICATION PENDING: phases 1–4 change generated output (fixed seeds
+produce different levels). Run SampleGenerator0 / AlienGridGenerator and the
+combat generator against pre-change captures before shipping. Phase 5 and the
+coverage pass are output-stable (the priority fix only affects nodes that were
+losing priority through transformer clones).