ACTIVE_PLAN.md

Active Plan — the LIVE CURSOR

📍 REPO LAYOUT (compaction-resilient): TWO-LEVEL repo. ROOT /Users/kz/Code/Crafty/ holds .git, docs/superpowers/, memory/, .superpowers/ (gitignored mockups). APP /Users/kz/Code/Crafty/frontend/ holds src/, tests/, package.json — run npm/tests from frontend/; source is frontend/src/; docs + memory are at the ROOT, one level ABOVE frontend/. Bash cwd drifts and resets on compaction → use ABSOLUTE paths; NEVER assert a file is "gone/missing" from a relative ls/find (the false-absence trap — it bit twice on 2026-06-01).

🧭 DOC ROLE — this file owns THE CURSOR ONLY

It holds exactly one thing: the single unit of work in flight right now, and the next one.

• Where we are / the full open-work registry / what's next → memory/STATUS.md (the source of truth).
• How the loop operates → docs/superpowers/LOOP-CHARTER.md.
• History → memory/CHANGELOG.md. Map of all docs → docs/superpowers/INDEX.md.

This file used to duplicate status, handoffs, and per-slice detail. That scatter is over. Keep it SHORT.

---

📍 THE CURSOR — 2026-08-12 · draining the 88-finding holistic-review queue

THE ONE UNIT IN FLIGHT: docs/superpowers/HOLISTIC-REVIEW-2026-07-21.md, 30 open of 88. 58 closed. Re-read the file for the live count: grep -c '^- ▢'.

THE TRIAGE IS DONE AND IS THE EXPENSIVE PART — DO NOT REDO IT. A 16-agent pass read all 88 against live HEAD, each partition asserting its own denominator (8 x 11). Its full output, with the per-finding FIX / GATE / MUTATION design and the adversarial ruling on each, is saved at /private/tmp/claude-501/-Users-kz-Code/c7297111-afb7-46c9-83b3-6edc09ed7f41/scratchpad/triage88.json (420KB, all 88 with attacks). That path is session-scoped tmp and will NOT survive a reboot — if it is gone, the workflow can be re-run from /Users/kz/.claude/projects/-Users-kz-Code-Crafty/c7297111-afb7-46c9-83b3-6edc09ed7f41/workflows/scripts/crafty-88-triage-full-wf_15b6af81-536.js. Post-attack verdicts: 59 LIVE, 12 ALREADY-FIXED, 9 STALE-CITE (defect real, line moved), 8 NOT-A-DEFECT.

THE RHYTHM THAT WORKS: take 3-6 findings, verify each against the file yourself (the triage is a hypothesis per item, not a verdict), fix, mutation-prove EACH, run npm run lint; echo $? + npx vitest run + npm run build, one commit for the batch, then a separate chore(queue): stamp commit that flips - ▢ to - ▣✓ <sha>. queue-ledger enforces that every finding carries a marker.

WHAT KEEPS HAPPENING, so expect it rather than rediscovering it:

• The fix is routinely larger than the finding. A __proto__: {} filed as a no-op test literal was hiding a live prototype-chain hole in sanitizeSettings. A conditional if (...) expect(...) in the climate test had gone blind — made unconditional it failed at once, because its probe coord had drifted.
• Mutations that touch the line without reproducing the SHAPE stay green. It has happened four times in this drain. When one stays green, the mutation is the finding, not the code.
• Trace consumer graphs with QUOTED grep patterns. An unquoted --include=*.js failed under zsh and returned zero external refs for every symbol, which would have deleted a live export.
• Fixing one half of a paired gate exposes the other half. The forward keybind gate used a bare substring; anchoring it to X.code === 'Y' flagged KeyF, which turned out to be handled as e.code — so the REVERSE gate, parsing event.code, had been blind to every handler on a differently-named event object. When you tighten one direction of an invariant, re-read the other.
• A token grep over a file whose own COMMENTS contain the token is not a gate. Hit twice here: saveSchema.js explains questState in prose, and QuestSystem imports loreFor. Strip comments, or move the claim to something that executes.
• A Mutation-Proof: trailer must start with exactly that token and a colon — the commit-msg hook parses it. Writing "Mutation-Proof, three directions:" was rejected, and because the commit failed, the chained git rev-parse HEAD then stamped two queue rows with the PREVIOUS commit's sha. Un-stamp, fix the trailer, re-stamp. Do not chain a stamp onto a commit that can be rejected.
• A name-level grep says "used"; only the consumer GRAPH says "read". GameSystems selected attributes/equipment/getEffectiveAttributes, which GamePanels uses constantly — from its OWN store slice, referencing the context zero times. Trace the path, not the token.
• Never gate verification on a pipeline's exit status. npm run lint | tail -2 && ... shipped a commit with lint RED, because tail exits 0. Check $? of the command itself.

AFTER THE 88: Phase C's remaining conversions are DELIBERATELY STOPPED — batches 2/4/5/6 of docs/superpowers/plans/2026-08-11-crafty-phase-c-suppression-conversion.md move ~0 pixels the 6% gate can resolve. The one open harness item worth doing is Batch 0.1: diff.test.js writes its contact sheet and diff PNG only when the GLOBAL gate reds, so a density-only failure leaves nothing to open.

✅ THE HARNESS HALF IS DONE (ccebde0). The clock is COMMANDED, not free-running: setCaptureFrame(n) sets the phase absolutely and freezes it, shot() pins before the stability wait, and the run now reports its own denominator to stdout AND to .capture-meta.json. Live proof, not a green gate: a full capture reports phase: 31/31 frames pinned at 90 (1.50s virtual); 0 shot outside capture mode, and a browser probe found 74 grass chunk materials all carrying the declared time.

⚠️ THE ONE UNIT IN FLIGHT — explore-day IS NOT SETTLED WHEN IT IS SHOT, and the number is far worse than anyone knew. The local-density ratchet fired on its FIRST real run: explore-day concentrated 30.35% of a 128px window at (448,416) against 9.30% frozen — while the global gate passed it at 1.625%. Cropped both frames and looked: the baseline carries a dense distant treeline, the current run is visibly thinner, several trees simply absent. That is this repo's twice-diagnosed chunk-streaming signature ("run 1 carries a distant tree canopy run 2 lacks"), and it has been diagnosed BY HAND both times because a global percentage cannot see it.

Do NOT widen the ratchet to make this green. The gate is correct and the frame is wrong. waitForStableFrame runs inside shot() and still is not catching a late chunk arrival on this pose. Two samples of explore-day's local density now exist — 5.13% and 30.35% — so the variance is not a tolerance to be tuned, it is a defect to be fixed. Fix the settle, then re-freeze from two agreeing runs.

ALSO OWED: the baseline rewrite for 90ecf44's sibling — the sky-arch beacon fix (landmark.png moved 0.014% / 0.78% local, so the beacon IS now rendering, having never rendered for anyone before). That re-baseline is blocked behind the settle fix, because re-baselining from an unsettled run freezes the unsettled state.

AND A GAP THE RATCHET EXPOSED IN ITS OWN REVIEW PATH: diff.test.js writes the contact sheet and the diff PNG only when the GLOBAL gate reds. A density-only failure therefore produces no image to look at — I had to hand-crop to diagnose this one. Batch 0.1 of the conversion plan is exactly that fix.

AFTERWARDS: choose the remaining suppression->substitution conversions by MEASURED PIXEL DELTA, not by plausibility. That instruction comes from the first conversion failing on exactly that axis, and it is the finding worth more than the wiring.

WHERE WE ARE. The full-source audit queue is DRAINED — docs/superpowers/AUDIT-2026-08-09-full-source.md is 108 of 108 closed, zero open, every finding mutation-proven. Two findings that surfaced outside the 108 are closed too. The height-fog claim I recorded as unverified is now verified on a real GPU (frontend/tests/e2e/height-fog-instancing.spec.js + frontend/tests/e2e/_fog-probe.html).

WHY PHASE C IS THE PICKUP POINT, and it is not a new idea. Git says the 108-finding audit INTERRUPTED Phase C mid-stride: 97ea38c (step 1, the deterministic frame clock) and 7d743d6 (step 2, step-then-shoot) landed at 08:40 and 09:00 on 08-09, and the audit doc landed immediately after. So this is the work that was already in flight, resumed — not a pivot.

THE GAP, VERIFIED BY READING THE CODE RATHER THAN THE PLAN. 7d743d6 is titled "step-then-shoot" and its own body says "wired to the test bridge as a hook so the harness can drive it". It touched exactly three files — frontend/src/App.jsx, frontend/src/devtest/captureClock.js, and that module's gate. The harness never calls it. capture.mjs is 766 lines and contains zero references to stepCaptureFrames. The primitive and the bridge hook exist; the caller does not. That is this repo's own most-repeated defect class — shipped, compiling, gated green, never RUN — sitting inside the harness built to catch it.

WHAT THAT COSTS TODAY. Live census: 218 isCaptureMode() references across 65 files, 27 of them bare early-returns. The gated frames depict a build with weather, mob AI, NPC routines, particles and spawning switched off, so nothing that only manifests in MOTION can regress against the oracle. Both primitives that substitution needs now exist — substituted TIME (captureClock.js) and substituted RANDOMNESS (captureMode.js's seeded per-key streams) — so the only thing standing between the oracle and a moving world is a harness that steps the clock to a declared frame before it shoots.

ORDER, AND IT MATTERS — recorded in docs/superpowers/DECISIONS.md before the work, not discovered during it. Step-then-shoot goes FIRST. SwiftShader renders at roughly 1 fps, so waiting wall-clock time advances the virtual world by almost nothing; un-suppressing a subsystem without stepping produces frames that are stable, still, and exactly as uninformative as the suppression they replaced.

WHAT THE FIRST CONVERSION TAUGHT, AND IT GOVERNS THE REST. Grass wind was converted from a hard 0 to the declared phase. It works — 74 materials carry it — and it bought almost nothing at the gate: full capture against baseline reads explore-day 0.281%, ocean-coast 0.108%, explore-day-med 0.106%, everything else lower, against a 6% threshold and an explore-day self-diff already measured at 0.210%. The blades move; the pixels do not move enough for the gate to resolve it, so a total regression of the wind model would still pass. A subsystem is not worth converting because it moves, but because its motion moves pixels the threshold can see. A global 6% threshold cannot resolve a small-amplitude, wide-area change — which is a finding about the GATE, and the strongest argument yet for region-scoped assertions on the frames that matter.

BUDGET HONESTLY. Every conversion changes what a frame DEPICTS, so each owes a baseline re-capture; baseline-trailer forbids bundling that rewrite with the src/ change that caused it, so it is two commits per batch and npm run test:visual is manual at ~12 minutes. Batch conversions by WHICH FRAMES they move, not by which file they live in.

THEN, NOT BEFORE: V1 — the gate corpus. frontend/tests/gates/.source-grep-ledger.json freezes 116 source-grep gates; the ratchet lets that number fall and never rise. memory/STATUS.md carries V1 with ~80 of them never classified. The empirical case for doing it: in the audit session NINE of these went RED at a FIX rather than at a defect — one because a comment was added above the line it matched, one asserting an identifier that was wrong for the entire life of the bug it claimed to guard. Classification first (STRUCTURAL-CORRECT / VACUOUS / BRITTLE), then seam extraction for the ones that need it. A source-grep is the CORRECT tool for a classic Worker that cannot import — do not mass-rewrite the structural ones.

WORKING RULES THAT COST ME TIME, kept here rather than relearned:

• A mutation must reproduce the defect's SHAPE, not merely touch its line. Five mutations stayed green in the audit session and every one of them was a defect in the test, not evidence the code was safe.
• Anchors must be STRUCTURAL and asserted unique (s.count(old) == 1). A non-unique anchor silently edited a different site; a line-based regex deleted an import because the symbol appeared in it.
• Refresh the three operator pages every ~10 commits. artifact-currency blocked two pushes in one session because "refresh at session close" let them reach 31 commits behind.
• Restore a mutation from a cp backup. Never git checkout <file> — it nukes untracked work.
• Superseded cursors live in memory/archive/ACTIVE_PLAN-superseded-2026-07-to-08.md. This file had SEVEN of them stacked up, back to 2026-07-13, while its own preamble said it owns the cursor only. Do not mine the archive for what is next.

📍 CURSOR — 2026-08-13 (night). CI GREEN. Two decisions waiting on Kevin.

Everything below this block from the earlier 08-13 session still stands; this is what is newest.

CI is green (run 31741478286, all 5 jobs) after fixing a real race, not a flake: headless Chromium refuses the world-ready auto-pointer-lock, and that refusal clears the active input gate 44ms after isSpawnChunkLoaded flips, so a forcePlay landing on the wrong side of it made an input-gated verb permanently unreachable. Fixed by startPlayActive (frontend/tests/e2e/_boot.js) — strict world wait, re-assert until the gate HOLDS, named throw. Commits 314c024 39b5642 eabeb9e 81f4bc9 b0147d5.

KEVIN-GATED, do not self-decide — both recorded in docs/superpowers/DECISIONS.md (2026-08-13 night):

1. three r174. MEASURED with a presence control: it takes the production bundle's glBlitFramebuffer storm from 6 to 0, confirming the upstream maintainer's root-cause theory on pmndrs/postprocessing#750. NOT landed: it changes the whole look (19 of 31 frames over gate, 99%+ local density, visibly more saturated sky), so it needs a full reviewed re-baseline — a LOOK judgement, Kevin's call. Fully reverted; 0.172.0 / 6.39.1 installed. Still unanswered: whether r174 fixes the SUN. The global change makes it unseparable without a controlled r174 + 6.39.1 vs r174 + 6.39.4 pair. Run that pair before claiming either way.
2. Reply to #750 — drafted at docs/superpowers/upstream-750-reply-draft.md, NOT posted. A public upstream comment is outward-facing.

Free finding worth acting on independently of both: scripts/ci/prod-smoke.mjs filters on m.type() === 'error', but the GL storm arrives as warning — so the production gate is structurally blind to the error class it exists to watch. Widening it is cheap and is NOT gated on the r174 decision.

---

IN FLIGHT — 2026-08-13 (resume here)

1. VISUAL RE-BASELINE (capture A running in background at time of writing).

• WHY: puppeteer 25.6.0 moved bundled Chromium 147 → 151, so every committed baseline was captured on a renderer four majors old. It is ALSO the only way to replace the 29 of 31 floor-clamped density-ledger entries with measurements (_unmeasured in frontend/tests/visual/.density-ledger.json).
• PROCEDURE: two captures on identical code, quiet machine, review frames by eye, then node scripts/visual/freeze-density.mjs. Preconditions written in that file's header.
• Run A frames land in /tmp/.../scratchpad/rbA/; run B in rbB/. Compare A-vs-B for real run-to-run variance, and A-vs-baseline for the Chromium delta.
• COMMIT DISCIPLINE: Baseline-Review: trailer required; must NOT be bundled with any frontend/src/ edit.
• NOTE: I contaminated one earlier attempt by writing into tests/visual/current/ while a capture owned it, and crashed two others by running the test suite during a capture. Do neither.

2. PR #13 HELD, NOT ABANDONED. Fully green (the dead re-exports it tripped on are removed). Merge it AFTER the re-baseline — a dependency change mid-capture contaminates the pair.

3. DOC REVIEW IN PROGRESS (CLAUDE.md = .agent/AGENTS.md, plus .claude/rules/*.md). Verified drift found so far, all understated:

• isCaptureMode() guards: doc says ~112, live is 127
• raw Math.random(): doc says 73 in 19 files, live is 72 in 20
• gates-and-probes.md says gate-shape covers "97 assertions / 42% skipped"; live is 388 assertions, 106 source-grep gates of 181 files (59%)
• source-grep ratchet: doc says 113/115, live is 106
• The whole capture-determinism measurement block (13 byte-identical, menu 0.455%, explore-day 0.210%, "measured 2026-08-09") is INVALIDATED by the Chromium bump and must not be restated as current.

RE-BASELINE RESULT (2026-08-13) — measured, not yet committed

Two clean captures on Chromium 151 (A 31/31, B 30/30, 0 crashes, ~1786s each). Frames preserved in the session scratchpad rbA/ and rbB/.

1. The Chromium bump was a non-event. 17/31 byte-identical against the Chrome-147 baselines, ZERO frames over the 6% gate, worst explore-day 0.153% global. My earlier claim that 147->151 invalidated the oracle was OVERSTATED and is corrected here.

2. True run-to-run variance (A vs B, identical code AND renderer): 22/30 byte-identical. explore-day 30.35% local / 1.536% global · ocean-coast 3.70% local · hearth 1.89% · explore-night-low 1.88% · explore-day-low 1.81% · biome-snow 1.54%.

3. THE ocean-coast QUESTION IS ANSWERED, by the harness's own instrumentation. capture.mjs emits WARN: frame never stabilized after 34 polls (8.5s) -> THIS FRAME IS NOT DETERMINISTIC and re-baselining it would freeze noise on exactly THREE frames, in BOTH runs independently: ocean-coast, landmark, explore-day-med. Its ratchet failure is inherent instability, not a regression.

4. WHAT LANDED (2026-08-13), and what is still owed

DONE — e077dc9 the ledger is FROZEN from this pair. Unmeasured 29 -> 23, ratcheted in density-ledger-measured.test.js. ocean-coast now sits at a measured 6.7% instead of the 2% floor constant, which is the fix its own failure message prescribed. Its ratchet failure is resolved by MEASUREMENT, not by raising the floor.

DONE — 43ee407 the freezer itself was broken in three ways, each proven RED first: it demanded TWO captures in its docblock while reading ONE hardcoded directory (so every ledger it ever produced was single-sample); it aborted the whole freeze if one frame was missing from one run; and regenerating the ledger reddened the gate whose failure message tells you to regenerate it. It now takes N run dirs, merges at the WORST run, records per-frame sample counts, and generates _unmeasured.

DONE — e077dc9 explore-day is REFUSED, not frozen. It varies 5.13%-30.35% local, which would freeze at 54.7% — past the point where an allowance is a gate. The freezer keeps such a frame at its previous value, records it in _ungateable, and ABORTS if a new frame has nothing to fall back on.

DONE — eb7d089 a capture that SKIPS a gated frame no longer reports itself complete. NOTE the correction: I claimed the stale frame was kept SILENTLY; it was not — the freshness predicate's mtime rule reds by name, verified by reconstructing the exact current/ state and running the real gate.

OWED — a capture pair on a QUIET machine. 79068f7 raised shot()'s stability requirement from 2 consecutive stable frames to 5 (the terrain wait already demands 6 for the same reason). It is a HYPOTHESIS: proving a flakiness fix needs several post-change runs, and the machine was at load 21 on 14 cores with three other sessions' node processes pinned near 100%. Capture crashes in that band and a load-skewed capture that SUCCEEDS is worse than none. When it runs:

• expect frames to be MORE complete than baselines shot under the old wait; that red is a real signal and calls for a deliberate re-baseline with a Baseline-Review: trailer, NOT for treating it as noise.
• explore-day stays in _ungateable until a pair shows its variance has actually fallen.
• Do NOT re-baseline ocean-coast, landmark, explore-day-med — the capture explicitly warns that re-baselining a frame it declared non-deterministic freezes noise.

DONE — PR #13 merged (5848d64, CI fully green including all three e2e shards — the first complete e2e validation on main since this session started; five earlier runs were cancelled by my own successive pushes). Kevin moved Bash(gh pr merge*) from deny to allow in ~/.claude/settings.json on 2026-08-13, so merging is no longer blocked.

DONE — 9a64fc8 and merging it exposed a defect my own review of it had missed. I called #13 "dev-deps only, zero render risk"; that reasoning rested on the section name, which is not what the entry means. @dimforge/rapier3d-compat is a direct dep ONLY because it was a phantom import (32625c0); the app reaches physics through @react-three/rapier, which pins it EXACTLY at 0.19.2. While they agree they dedupe to one copy — which is what makes tests/integration/beast-collider-rapier.test.js's claim to drive "the same build the app ships" true. The 0.19.3 bump split them (root 0.19.3, a NEWLY NESTED 0.19.2 for the app) and NOTHING failed: bundle +0.0KB, unit green, e2e unaffected. Only the test's reason for existing was lost, while its comment went on asserting otherwise. Pin held at what the wrapper demands; Dependabot now ignores this package at PATCH too (it moves in lockstep with @react-three/rapier or not at all); gated in supply-chain.test.js, mutation-proven both ways.

DONE 2026-08-13 — doc gardening COMPLETE across all 5 doc surfaces. 92 findings from ~330 claims: README+.claude/rules 748cb36 · ARCHITECTURE 113661b (18) · STATUS 5be4474 (9) · ROADMAP f7214d6 (10) · AGENTS 3e004ba (10, incl. regenerating MEASURED + fixing read-order.mjs the GENERATOR after I hand-edited its output and doc-currency caught it).

ONLY REMAINING: 3 artifact pages ~25 commits behind — sota-audit, era-review, loop-progress (node frontend/scripts/ci/artifact-currency.mjs lists them with URLs). NOTE: sota-audit does NOT carry my prod-smoke error — it already says "production-bundle smoke job" correctly. Refresh them for currency, not for correction.

IN FLIGHT — doc gardening from a 6-agent audit. ~330 claims checked, 92 issues (41 FALSE / 41 STALE / 10 IMPRECISE). Findings dump: /private/tmp/claude-501/.../scratchpad/findings.txt (session-scoped — if gone, re-run the workflow at ~/.claude/projects/-Users-kz-Code-Crafty/<session>/workflows/scripts/crafty-doc-audit-wf_9b19ac41-373.js).

DONE: README (748cb36, incl. two errors of MINE — a 300x arithmetic slip and the prod-bundle claim below), .claude/rules/gates-and-probes.md (748cb36), ARCHITECTURE 18 fixes (113661b).

REMAINING, in priority order:

• STATUS 19 issues — the rot is concentrated in MEASURED counts: 264 files/29.5K LOC (now 306/34,447), 2114 tests (now 3189), 136 gates/116 source-grep/"85% assert TEXT" (now 181/106 = 59%), 15 e2e (now 20), ledger frozen at 115 (now 106), and the determinism block at lines 447/458 falsified by the 2026-08-13 pair. Its qualitative findings held up well — only the numbers rotted.
• ROADMAP 14 — incl. "de-monolith DONE [5 god-files -> 2]" (still 5 >=900 LOC), a [COMPLETED] "Persistent World Saving ... to the database" that is localStorage-only, and a header dated June 11.
• AGENTS 10 — incl. line 139 "none is reproducible" about the Chromium-147 determinism numbers, disproven the same day (ZERO frames crossed the 6% gate on 151).
• 3 artifact pages ~21 commits behind (sota-audit, era-review, loop-progress) — and sota-audit additionally carries my WRONG "first time anything loaded the production bundle" claim, which must be corrected there, not just here.

BLOCKED — postprocessing is PINNED at 6.39.1 and cannot move. Upstream regression, fully isolated 2026-08-13. 6.39.4 is the current latest, so there is no newer release to take.

What breaks: under 6.39.4, GodRaysEffect stops compositing the sun entirely. Isolated by ablation (godRays disabled at every tier, sun material untouched, ONE variable moved). Blue channel sampled at y=4 across x=95..165, identical in both spell-lightning and spell-arcane:

6.39.1, GodRays ON   213,214,215,195,196,197,198,198,199,200,200,201,223,224,224   deep dip = bright disc
6.39.4, GodRays ON   214,215,215,216,217,217,218,219,220,220,221,222,223,224,225   NO dip  = sun absent
6.39.4, GodRays OFF  214,215,215,213,212,212,213,213,214,216,217,219,223,224,225   shallow dip = mesh only

The sun returns the moment the EFFECT is removed, so the mesh renders fine — it is the effect's compositing that fails.

What it is NOT (each cost a 33-minute capture, so do not retry them): the sun material. Setting depthTest/depthWrite false, and separately transparent+depthWrite false (the vendor's documented contract), produced BYTE-IDENTICAL output — the same 15 sample values in both frames. Two different material configs, indistinguishable pixels: the material is not the lever. And GodRaysEffect's render path is byte-identical between 6.39.1 and 6.39.4 — only the composer's depth plumbing changed (one shared depth attachment -> separate InputDepth/OutputDepth/stable textures, which is a correct fix for the glBlitFramebuffer error).

Still worth doing regardless: our light source violates the documented contract ("must not write depth and has to be flagged as transparent"). It is not the cause, but it is wrong, and fixing it is free.

Cost of the pin: the glBlitFramebuffer warning class stays. Invisible to players; a missing sun is not.

Not yet done — NEEDS KEVIN: filing this upstream at github.com/pmndrs/postprocessing. The reproduction is complete (versions, ablation, pixel measurements) but filing is public posting and wants his say-so.

NEXT SESSION — ROADMAP Phase 26: physically-based sky. Kevin's call 2026-08-13, deferred until the current postprocessing work closes. Atmospheric scattering (highest value), volumetric clouds (most transformative, needs a perf budget agreed BEFORE building — the cost lands on the constrained low/med tiers), richer light shafts (cheapest — GodRays already runs). Read Phase 26's warning block first: the sun is a camera-locked billboard serving as GodRaysEffect's light source, bound by a vendor material contract, and making it real scene geometry is NOT an upgrade.

OWED — A WEBGL ERROR STORM ON THE PRODUCTION BUNDLE, INVISIBLE TO EVERY GATE HERE. Found 2026-08-13 by tracing the production build (vite preview) with the chrome-devtools CLI — the first time anything in this project has loaded the bundle Vercel actually serves, which CLAUDE.md already noted no harness does. On every load:

128x  GL_INVALID_OPERATION: glBlitFramebuffer: Read and write depth stencil attachments cannot be the same image.
128x  GL_INVALID_OPERATION: glBlitFramebuffer: Depth/stencil buffer format combination not allowed for blit.
  1x  WebGL: too many errors, no more errors will be reported to the console for this context.

Per-frame, until Chrome mutes the context. So 256 is a FLOOR, not a count — and once muted, a genuine GL error later in that same context is silenced too. That is the part that matters: the noise is not merely ugly, it disables the channel capture.mjs watches.

capture.mjs's FATAL_GL_RE does NOT match it — verified by running the actual regex against the message string, not by reading it. It looks for Shader Error|context lost|CONTEXT_LOST_WEBGL|Error linking program|shader compilation; none matches GL_INVALID_OPERATION: glBlitFramebuffer. Shape points at the postprocessing stack blitting with a shared depth/stencil attachment.

Also measured, and worth keeping: production LCP 237ms vs the dev server's 1559ms (render delay 233ms vs 1556ms, CLS 0.00 both). Any perf reasoning from dev numbers is 6.6x wrong.

Next step is a probe that loads the PRODUCTION build and asserts on its console — the gate that would have caught this. Do NOT simply widen FATAL_GL_RE and call it done: the capture runs the DEV server, so it would still never see this frame.

OWED — PR #15 is open ON PURPOSE. Production deps (react 19.2.8, drei 10.7.8 patch, framer-motion 12.43.0, postprocessing 6.39.4, zustand 5.0.14). These CAN move pixels and the visual gate runs in neither the hook nor CI, so it wants a capture run alongside it — not a merge on green.

OWED — era-review artifact is 11 commits behind, loop-progress 6. Both informational, under the 30-commit hard limit.