feat(sim,render): #128 underground-embedding guards (PR 6-sim V30) + sprite containment (PR 6-render)#208
Open
LightAxe wants to merge 15 commits into
Open
feat(sim,render): #128 underground-embedding guards (PR 6-sim V30) + sprite containment (PR 6-render)#208LightAxe wants to merge 15 commits into
LightAxe wants to merge 15 commits into
Conversation
…28) — flurry PR 4 PR 4 of the #127/#128 fix series (plan/flurry/PR2-FIX-SPECS.md). Freezes surface terrain so it never mutates as food piles spawn/deplete or entrances are designated — the static-terrain bug class Phase 0 measured (477/900 piles revealed a HardBlock on removal). Scope is STRICTLY static terrain; no #127/#128 movement fixes (those are PR 5 / PR 6). What changed: - surface-features.ts: deleted the dynamic entrance/food-pile suppression from the shared feature selector (movement AND render). Split procedural vs frozen: surfaceFeatureProcedural (terrainSeed only) + a baked 128x128 movement-effect grid that is the source of truth for surfaceMovementAt (O(1)). surfaceFeatureAt consults the carve override so render never paints a sprite over a carved-passable tile (R4-3). Added bakeStaticTerrain (deterministic root-clearance reservation + corridor carve guaranteeing ONE connected component; no terrainSeed retry), computeSurfaceComponentMask, isSurfaceTileInComponent, validateSurfaceConnectivity. Boot-asserts no registry feature is Cosmetic (carve-detection precondition). - types.ts: SIM_VERSION_V28_STATIC_TERRAIN (LATEST=28); new stored Uint8Array bakedSurfaceEffect + derived surfaceComponentMask; procedural bake in createWorldState; threaded copyWorldState. - save.ts: bakedSurfaceEffect serialized packed 2-bit + base64 (~5.3 KB vs ~48 KB raw). Deserialize validates dims / enum range / full connectivity (every saved pile + entrance), rejecting corrupt or pre-V28 maps. MIN_ACCEPTED raised to V28 (posture 2 — old saves reject cleanly). - scenario.ts: bake BEFORE placing piles; reachable-spawn pile gate; world-gen connectivity assertion. - tick.ts: DesignateEntrance rejects unless candidate + clearance are already passable on the frozen grid AND in the connected component (can no longer carve). - food-system.ts: runtime pile spawn gated on the connected component. - constants.ts: SURFACE_ROOT_CLEARANCE_RADIUS. Oracle/tests: featureFieldHash extended to the FULL SurfaceFeatureSlice (kind+variant+anchor); committed oracle (0 mutate across pile spawn/deplete, entrance designate/open, save/load), connectivity, and simVersion rejection-boundary tests; determinism serializer threads the baked grid; suppression/render tests updated to the static-terrain contract. Measurements: save-size delta +0.49%; tick-time ~0.24 ms/tick; acceptance sweep (10 seeds x 3 difficulties) 0 connectivity failures, 0 hash mutations. verify green (79 files / 2336 tests). Does NOT close #127/#128 (PR 5/6). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…view) Addresses the PR #206 CI failure + Codex/ship-review findings. - Perf (CI timeout fix): createWorldState + createScenario ran a full 16,384-tile procedural bake on every call (twice per scenario), tipping many-iteration tests over the 5s timeout on the CI runner. Memoize bakeSurfaceEffectGrid by terrainSeed (pure function of it) and return a copy; the suite is now faster than the pre-PR baseline (tests ~31s vs ~60s). - save.ts base64 (Codex P2): reject `=` padding outside the final quartet (and the never-valid "x=y" form) instead of treating it as zero — malformed encodings now fail the load boundary as intended. - SAVE_FORMAT_VERSION 3 → 4 (ship-review): the new required on-disk bakedSurfaceEffect field is a shape break; bump it (and SAVE_KEY :v3→:v4 + purge old key) so save-shape correctness isn't silently coupled to the simVersion floor. - DesignateEntrance halo (ship-review, bugs dim): check component membership per clearance tile instead of a bare HardBlock test, so it no longer relies on an unenforced "globally one walkable component" assumption (isolated empty pockets are harmless). Softened the over-claiming doc comments to match what the bake actually guarantees (all roots/piles/entrances share one component). - computeSurfaceComponentMask: throw on a wrong-sized grid (fail loud instead of silently flooding the map into one component). verify green (79 files / 2336 tests). Still PR 4 scope only; does NOT close #127/#128. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rio (codex P1) Codex P1: the procedural-bake `Map` was mutable module-level state in src/sim — an AGENTS.md ECS-rule blocker (state outside the world snapshot). Removed it. Restore the performance without a cache: - bakeStaticTerrain now carves a COPY of the procedural grid createWorldState already baked, instead of re-baking from scratch — one full procedural pass per createScenario (was two). Measured: 200x createScenario ~2.0s (was ~4s). - vitest.config: raise the default testTimeout to 15s. Static terrain made world construction bake the 128x128 grid (~8 ms each), so construction-heavy sweeps (200-seed spider-lair, 500-iteration pause-cadence) sat near the 5s default and tipped over it under CI-runner contention — the PR #206 CI timeout. Assertions and iteration counts are unchanged; genuinely-hung tests still fail. verify green (79 files / 2336 tests, ~57s). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ut fix The per-tile procedural bake called the recursive overlap-suppression check up to 16,384× per grid (~8 ms), so construction-heavy tests (200-seed spider-lair, 500-iteration pause-cadence) blew even a 15 s timeout on the CI runner. Raising the timeout only masked it; this fixes the root cause. bakeSurfaceEffectGrid now iterates ANCHORS in lex (ay,ax) order instead of tiles: the first surviving anchor covering a tile claims it (matching the per-tile lex-smallest-winner tie-break), with a per-anchor "first covering feature per offset" scratch to reproduce the selector's first-covering-feature + break-on- suppress semantics exactly, including negative (off-edge) anchors. Verified byte-identical to the per-tile surfaceFeatureProcedural across 80 seeds × all 16,384 tiles (0 mismatches), so surfaceFeatureAt's carve detection and the feature-field oracle stay consistent. Measured: ~0.55 ms/bake (was ~8 ms); 200× createScenario ~0.45 s (was ~2-4 s). Also: - Reverted the vitest testTimeout bump — no longer needed; the bake is fast. - createScenario resets surfaceComponentMask after initColony so the persisted/ validated mask is rooted at the real canonical root (first colony entrance) rather than relying on the pre-colony fallback root happening to equal it (ship-review advisory). verify green (79 files / 2336 tests). PR 4 scope only; does NOT close #127/#128. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…(codex P2) The DesignateEntrance sim gate (PR 4) now rejects tiles outside the connected walkable component or with a blocked clearance halo, but surface-input still previewed any empty tile — so a player saw a valid-looking entrance target and confirmation did nothing. Add isValidEntranceTarget (empty tile AND candidate + SURFACE_ROOT_CLEARANCE_RADIUS halo all in-component) mirroring the gate, and use it for the right-click preview. Smoke tests updated: the mock world carries a fully-walkable baked grid, plus new cases asserting no preview when the target or a halo tile is HardBlock. verify green (79 files / 2338 tests). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ia simVersion The e2e CI job failed (5 save/load specs): bumping SAVE_KEY v3→v4 orphaned every spec that addresses the literal 'subterrans:save:v3', and the bump itself was only a low-severity ship-review advisory, not a spec requirement. Revert SAVE_FORMAT_VERSION (4→3) and SAVE_KEY (:v4→:v3). The new required bakedSurfaceEffect lives in the SNAPSHOT (SerializedWorldState), and this codebase versions snapshot-content shape changes by simVersion — not the envelope's SAVE_FORMAT_VERSION (which versions envelope structure). deserializeWorldState validates simVersion FIRST, so raising MIN_ACCEPTED_SIM_VERSION to V28 already rejects every pre-static-terrain save (which also lacks the field) before unpack is reached — the required behaviour, achieved without an envelope bump and matching the v2/v3 precedent. The legacy-purge list reverts accordingly. All required PR-4 save behaviour is retained: MIN_ACCEPTED=V28, the required bakedSurfaceEffect field, and load-time dims/enum/connectivity validation. Verified locally: npm run verify green (2338); the 5 previously-failing e2e specs (menu-and-dialog #115/#196, phase-09 save-prompt) now pass (19/19 chromium). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ailing-bit guard Two low-severity ship-review advisories on the latest diff: - surface-features.ts: the @param world JSDoc on isAnchorSuppressedByOverlap still described the deleted isAnchorGameplaySuppressed call. Correct it — world is now only threaded to the recursive calls; no field is read (PR 4 removed dynamic suppression). Doc-only. - save.ts: unpackBakedSurfaceEffect now rejects non-zero unused high bits in the final packed byte when expectedLen isn't a multiple of 4, so a tampered payload can't decode to a grid serialize would never emit. No-op for the production 16384-tile grid (a multiple of 4), but keeps the exported, length-general validator honest. (Left as-is: the preview/sim entrance gates still diverge on column-uniqueness / entrance-cap — that divergence predates PR 4; this PR only added and mirrored the new component/clearance gate. Out of scope.) verify green (79 files / 2338 tests). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…lence/serialization tests, hardening Fable (adversarial review standing in for Codex) found no runtime bugs but flagged spec-mandated validation/test gaps + low-severity hardening. All addressed: P2 — load-validation + load-bearing test coverage: - validateSurfaceConnectivity now also checks each entrance's full radius-3 clearance halo is in-component (spec R3-8), mirroring the DesignateEntrance gate — so a corrupt/tampered save the sim gate could never produce is rejected. - Committed the bake↔selector equivalence sweep (bakeSurfaceEffectGrid byte- identical to per-tile surfaceFeatureProcedural across 6 seeds × all 16,384 tiles), plus direct computeSurfaceComponentMask and corridor-carve (carveCorridor merges a HardBlock-split row) regression tests. - §6.5 field-specific serialization tests for bakedSurfaceEffect: create→copy round-trip (determinism.test), save→load mutate-and-survive (catches a re-derive-from-seed deserializer), every unpackBakedSurfaceEffect rejection path, and load-time connectivity rejection. P3 — robustness: - Oracle entrance-designation step asserts ≥1 entrance was actually accepted (no vacuous rejection-only pass). - surfaceComponentMask materialized eagerly at world-gen, load, AND tick entry, so input/render queries (and mid-tick spawn checks) are pure reads, never a lazy world mutation. - isValidEntranceTarget halo bounds use the grid CONSTANTS (matches the sim gate). - packBakedSurfaceEffect throws on out-of-range effect values (fail at write, not a later read); unpackBakedSurfaceEffect rejects by base64 length BEFORE decoding (DoS guard, matching the issue #99 load-cap posture). - validateSurfaceConnectivity rejects the degenerate all-HardBlock-root case (zero entrances + zero piles no longer passes on a garbage all-zero mask). verify green (79 files / 2345 tests, +9 new); internal ship-review passed:true. Still PR 4 scope (static terrain); does NOT close #127/#128. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
Owner
Author
🔍 Adversarial Review (Claude, standing in for Codex)Reviewed commit: What I verified before the findings: the tightening-write enumeration is complete — I walked every P2 — The queen has a second, unguarded descent path (class-ii guard incomplete) The class-ii landing guard was added to the worker descent ( P3 — Class-ii structural coverage is worker-only The committed class-ii test descends a CarryingFood forager onto a non-enterable landing. Neither queen descent path has a structural test — add one alongside the 3686 fix (a queen on an open entrance with a non-enterable column top must stay on the surface), so both queen sites are pinned against regression. VERDICT: REVISE — everything else is clean: the guard set is provably complete, the transactional semantics are right, the dead-digger lifecycle can't orphan, and the render invariant is both proven and non-vacuously probed. The single P2 is a four-line fix plus one test. |
Owner
Author
🔍 Adversarial Re-Review (Claude, standing in for Codex) — round 2Reviewed commit: Both round-1 findings are addressed — verified against the diff:
Also noted: For the post-rebase ledger, the only item still open anywhere in the series is the non-blocking PR 4 nit (length pre-check shadows the decoder rejection-test labels in VERDICT: APPROVED — all three PRs in the series are now approved at their current heads (#206 |
…te (Fable P3) The malformed-input cases in the unpackBakedSurfaceEffect rejection test were all shorter than the expected base64 length for a full grid, so the DoS length pre-check rejected every one of them before the decoder branch each label named could run. Rebuild each case at the exact full-grid length (5464 chars) with a single targeted corruption — bad char, mid-string padding, x=y final-quartet padding — so each exercises the base64ToBytes branch it claims, and keep an explicit wrong-length case for the length gate itself. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
… (C-both) — flurry PR 5 (V29) PR 5 of the #127/#128 fix series, built on PR 4 static terrain. Eliminates the dominant #127 mechanism (scent/priority-vs-wall, 86% of episodes) and the residual wander short-tail. Fix-A — passability-aware stepping (surface-routing.ts): - New stepTowardReachable(world, from, target): first step of a COMPLETE BFS goal field (distance-to-target over PR 4's single connected component), cached per target tile on the world (derived, not serialized). Replaces the naive pickCardinalStep in BOTH forager target branches (priority + scent). Target identity unchanged — only the step is path-aware. Three-valued: AtGoal (0,0) / Step / InvariantViolation. Scent ranks Manhattan-15-eligible piles by reachable path distance, lowest-foodPileId tie-break. No-revisit bypassed for targeted steps. Zone-guarded to the surface (underground falls back to cardinal step). Diagonals chosen only when a shared orthogonal is itself passable+descending (no corner-squeeze that the movement guard would reject). C-both — recent-tiles deepening: - RECENT_TILES_LEN 4→12 (chosen N: largest of {12,10,8,6,5} meeting all caps). - Compact canonical save encoding (records sorted by antId; head + non-sentinel (slot,x,y) in slot order); load validates + rejects malformed streams. Far smaller than the old flat arrays — total save size −19% vs baseline. simVersion V29 (posture 2): bump LATEST + raise MIN_ACCEPTED. Pocket-escape not needed — deepening alone drives confinement to zero. Acceptance (committed harness tests, ACCEPTANCE hold-out): aimedIntoWall=0, worst confinement=0tk (≤60), episodes>300=0; 4-vs-N sweep N=12 meets every throughput cap; field-specific copy/save-load + V29 boundary tests; save −18.95% (≤+5%); tick-time 0.26 ms/tick (≤0.5). verify green (80 files / 2380 tests); ship-review passed:true. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
… metric, honest zone comments, ring-load hardening Fable (adversarial review standing in for Codex) flagged that the acceptance story had drifted. All findings addressed: P2 — acceptance integrity: - harness stepLandsOnWall reverted to the NAIVE pickCardinalStep direction (the #127 wall-aim signature), measured INDEPENDENTLY of the router. Calling stepTowardReachable made the cap tautologically 0 (the router never returns a wall step). Fix-A satisfies the cap by eliminating the confinement episodes the flag is tallied within (episodeAimedWall is only set during a confined run), not by redefining the flag — now falsifiable again. - The underground scent pull is now DELIBERATELY dropped in V29 with an honest comment (it fed SURFACE pile coordinates to an underground walker — meaningless; underground searchers use pheromone/wander). The prior "(pre-PR-5 behaviour)" comment was wrong. - The unreachable-surface priority case is documented as the benign transient stale-target handling it is (PR 4 guarantees current priority piles are reachable, so an unreachable target is a leftover from a prior Fighting/zone stint). Investigated alternatives: asserting crashes on these (S3 determinism scenarios); falling through to wander REGRESSES worst confinement to 108 tk. The cardinal step + surface-detour guard keeps the ant moving (harness confirms worst=0, aimedIntoWall=0), so it is kept — stepTowardReachable's throw remains a defensive guard. P3: - packRecentTiles comment corrected (antId<nextEntityId skips NEVER-allocated ids; dead-but-allocated ARE included). - unpackRecentTiles now rejects antId outside [0, min(nextEntityId, capacity)) — nextEntityId is the canonical (serializer) bound, capacity the memory-safety bound; a tampered count/nextEntityId mismatch is rejected, not silently dropped. - measureForagingThroughput doc: underground-carrier behaviour is subsumed by the cross-zone delivery latency/path/completion counters. - Added a goal-field cache-eviction (clear-all-on-overflow) regression test. verify green (80 files / 2381 tests); ship-review passed:true. Still PR 5 scope (V29 path-aware routing + deepened recent-tiles). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Owner
Author
|
@codex review |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
…review round 1) Entity ids are never reused and dead rings are never read (all isRecentTile/detour consumers run only for alive ants), but an ant killed mid-Searching/CarryingFood keeps its populated ring in memory forever — serializing those records grew every subsequent save monotonically. Skip non-alive ids at pack time; old saves containing dead-ant records still load (unpack doesn't check alive) and self-heal on the next save. Tests now seed rings on live worker ids instead of raw ids 0/1/3, plus a regression test that a dead ant's populated ring is not serialized. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…sprite containment (PR 6-render) PR 6 of the #127/#128 fix series, built on PR 5. Closes the latent #128 underground-embedding invariant via two artifacts. PR 6-sim (V30, posture 2 — bump LATEST + raise MIN_ACCEPTED): - (ii) Descent landing-tile guard (worker + queen): an ant lands underground ONLY on a tile it canEnterUndergroundTile for its task; else it holds on the surface (re-checked next tick) — never seeks a nearest legal landing. Fails CLOSED if the grid is absent. - (iv) One task-aware occupancy guard on EVERY passability-tightening underground mutation, in the neutral underground-occupancy.ts module (findEmbeddedByTightening, keyed by currentGridColonyId — sees foreign occupants — placed neutrally to avoid the colony↔ant-system import cycle): - CancelDigMark Marked→Solid: single-tile guard; chamber-cancel guarded TRANSACTIONALLY (preflight the whole footprint, block-all-or-revert-all). - tickDeadDiggerCleanup BeingDug→Marked: retain-the-dead-claim-and-retry under a non-digger occupant (never orphans the BeingDug tile); liveness bound documented. - PlaceChamber Solid→Marked is loosening — not gated. PR 6-render (no simVersion bump — render-only): - sprite-containment.ts: containedScale() uniformly scales an underground ant sprite down just enough that its rotated AABB never crosses its Open anchor tile's boundary toward a Solid/off-grid neighbour (overflow into passable neighbours stays allowed). Wired into draw-underground.ts. The queen (20×14) and rotated fighters overflow the 16px tile pre-fix; workers at natural size do not. Acceptance (committed harness/probe tests, all printed): - #128 structural (ii descent + iv CancelDigMark/chamber-cancel/dead-digger incl. orphan-prevention) = 0 embed; adversarial dense-cancel stress = 0; AI record+replay (runAIController) = 0 natural embeddings. - Render containment probe: 525 configs (kinds × rotation extrema × max scale × interp-alpha sub-tile centers) — 0 dirt overlaps; non-vacuous (unclamped queen overflows) and does not over-clamp on passable neighbours. - V30 rejection-boundary; determinism + neutrality green; save −18.95% (≤+5%); tick-time 0.25 ms/tick (≤0.5). verify green (81 files / 2390 tests); ship-review passed:true (2 low advisories proactively addressed: fail-closed descent, dead-digger liveness bound). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…lass-ii P2) Fable found the V30 class-ii landing-tile guard was on 2 of 3 descent sites: the worker descent and the queen's step-onto-entrance descent were guarded, but the queen's PRE-MOVE descent short-circuit (already standing on her open entrance, ant-system.ts ~3686) still set zone=Underground/posY=0 with no enterability check — the same class-ii embed, reachable via the documented starter case or a corrupt save (isOpen true + non-enterable column top). - Add the same fail-closed landing-tile guard to the pre-move queen descent. - Add a structural test: a queen on an open entrance with a non-enterable column top stays on the surface (pins all queen descent paths against regression). - Minor: un-export rotatedAabbHalfExtents (internal-only; review surface nit). verify green (81 files / 2391 tests); ship-review passed:true (2 low advisories: documented diagonal-clamp conservatism; aabbOverlapsDirt kept as the committed containment-invariant predicate the probe asserts). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ntainedScale (ship-review round 1) Two over-clamps drove the contained sprite scale to 0 in normal play: - isDirt treated every off-grid tile as dirt, so a freshly-descended ant (posY=0, sprite center on the grid's top edge, edgeN=0) clamped to scale 0 and popped out of existence on every descent. Nothing renders above row 0 in the underground viewport, so the above-grid row is now exempt. - A Solid diagonal clamped BOTH axes to that corner's edge distances unconditionally, collapsing the factor toward 0 whenever the center neared a tile edge with dirt diagonally ahead — every tile crossing in a 1-wide tunnel. A corner only obstructs when the AABB crosses both of its edges, and shrinking along one axis suffices, so it now clamps only in that case and only along the less-restrictive axis. Both keep the containment invariant (footprint never paints dirt); regression tests fail against the previous clamping. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
LightAxe
force-pushed
the
flurry-pr6-embedding
branch
from
2187609 to
16d17fe
Compare
Owner
Author
🔍 Adversarial Re-Review (Claude, standing in for Codex) — round 3 (post-rebase)Reviewed commit: The two containment over-clamp fixes are correct, and I checked the geometry adversarially rather than taking the commit message's word:
Both fixes ship with regression tests that fail against the previous clamping, and the original probe suite (vacuity control, no-over-clamp, worker-never-clamped) still passes around them. No findings. VERDICT: APPROVED (head |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR 6 of the #127/#128 fix series — closes the latent #128 underground-embedding invariant. Built on PR 5 (#207) → stacked on #207, which is stacked on #206; review/merge #206 → #207 → #206 first. simVersion V30 (PR 6-sim); PR 6-render is render-only (no bump). Do not merge until Rob OKs.
PR 6-sim (
src/sim, V30, posture 2)canEnterUndergroundTilefor its task, else holds on the surface — never seeks a "nearest legal landing". Fails closed if the grid is absent.underground-occupancy.ts(findEmbeddedByTightening, keyed bycurrentGridColonyIdso foreign occupants are seen; placed neutrally to avoid the colony↔ant-system cycle):CancelDigMarkMarked→Solid — single-tile guard; chamber-cancel guarded transactionally (preflight whole footprint, block-all-or-revert-all).tickDeadDiggerCleanupBeingDug→Marked — retain-claim-and-retry under a non-digger occupant (never orphans the BeingDug tile); liveness bound documented.PR 6-render (
src/render, no bump)sprite-containment.ts—containedScale()uniformly scales an underground ant sprite down just enough that its rotated AABB never crosses its Open anchor tile's boundary toward a Solid/off-grid neighbour (overflow into passable neighbours stays allowed). The queen (20×14) and rotated fighters overflow the 16px tile pre-fix; workers don't. Wired intodraw-underground.ts.Acceptance (committed tests, all printed)
Verification
npm run verifygreen — 81 files / 2390 tests; internalship-reviewpassed:true(2 low advisories proactively addressed). Codex review pending (account out of credits); reviewed internally + by Fable.🤖 Generated with Claude Code