fix(templates): share backlight graph and honor the render camera

biome.json

@@ -43,6 +43,7 @@
       "examples/csg-doorway/public/**",
       "docs/benchmark/sweeps/**",
       "docs/benchmark/rounds/**",
+      "docs/benchmark/prd345/shooter-causes1/**",
       "docs/verification/prd-357-c4-duplicate-inventory.json",
       "docs/verification/prd-357-c4-duplicate-inventory-before.json",
       "docs/verification/visuals/**",

docs/PRDs/rendering/PRD-345-a-backlit-subject-is-not-a-hole-in-the-sky.md

@@ -4,7 +4,7 @@ prd_contract: v1
 
 # PRD-345 — a backlit subject is not a hole in the sky
 
-**Status:** PARTIAL — generated material defaults integrated across conventional templates; twelve final matched web desktop pairs are palette-stable, including corrected racing. All-template gameplay/full-root/native appearance and rain admission remain open. Originally filed 2026-09-03, measured at `43d03e6a`. Batch:
+**Status:** PARTIAL — generated material defaults integrated across conventional templates; twelve final matched web desktop pairs are palette-stable, including corrected racing. A controller-owned shared render graph removes the shooter's per-material-graph performance regression in the isolated causal study (`docs/benchmark/prd345/shooter-causes1/`); integrated gameplay/performance re-qualification, all-template/full-root/native appearance and rain admission remain open. Originally filed 2026-09-03, measured at `43d03e6a`. Batch:
 [docs/PRDs/AAA-visuals](aaa-visuals-notes.md). **Ships as generated user source, not as a package** — it
 decides how things look, and rule 1(b) vetoes 1(a) at any size. Source studied:
 [TheLongSilence](https://github.com/achimala/TheLongSilence) `src/gfx/greeble.js:37`, the
@@ -97,12 +97,18 @@ P1, PRD-345. The workbook remains unchanged.
 ### Phase 2 — admit measured generated defaults
 
 - [ ] Apply only qualified conventions to the actual starter/material source and all required templates, preserving per-template appearance, low/mobile/software fallbacks and named overrides. proof: matched per-template screenshots and actual frame/startup cost
+  - 2026-10-04: the twelve material templates now build **one controller-owned render graph** (`src/render/backlightControls.ts` owns `validateControls`/`resolveKey`; `backlightMaterial.ts` caches one graph per controls object in a `WeakMap` so every converted material references the same graph) instead of a per-material graph. The isolated 21-launch causal study at the unchanged 1920×1080 / 33 ms limit measured per-material current 46.93 ms (0/6 pass) against shared graph 30.43 ms (3/3 pass), with copy-only (29.37 ms) and material-emissive-only (28.90 ms) controls also passing and runtime-zero/compiled-rim-only controls failing. Evidence: `docs/benchmark/prd345/shooter-causes1/` (raw reports, source identities, PNG hashes). Integrated gameplay/performance re-qualification, the snow threshold and native remain open, so this box stays open.
 - [x] Add the concise convention/override rows to generated instructions with mirrors kept in sync. proof: all thirteen template rows, explicit rain exception, primary-docs/instruction-budget seventeen tests and scaffold/doc eighty-three tests pass; unchanged instruction caps
 
+2026-10-05 PR #440 merged-source checkpoint (`develop` 45565868): independent review identified a second-camera layer error, reproduced by a regression test in the shared key callbacks. The callbacks now pass actual `NodeFrame.camera` to key resolution in all twelve material templates and the matching fixture. The existing test failed before the fix and the camera/matrix/lifecycle suite passes 31/31; merged scaffold/compiler contracts pass 101/101 with all thirteen hashes measured through the real generator. The ordinary JS/DTS build passes. Fresh serial hardware NVIDIA Turing WebGPU difficult-lighting captures pass the unchanged decoded-PNG qualifier: enabled edge/body p99 52/2 versus zero-rim 7/0; fill-black fails only the two intended tone assertions, and omitted reporting fails the marker verifier. Actual camera world/inverse/projection matrices are now recorded. Local evidence is retained at `artifacts/pr440-acceptance/`; images remain unpublished pending the user's sharing decision. Failed startup wrappers and one unobserved black-fill launch remain retained beside its observed retry. This verifies the corrected graph's isolated controls; integrated template gameplay/performance, repeated startup and native admission remain open.
+
+2026-10-05 integrated checkpoint: all six fresh unchanged 33 ms runs are **RED**: shooter p95 112.9/144.5/88.9 ms, snow 59.8/48.6/51.8 ms; one shooter also fails minimum FPS. Host/resource and source differences prevent code-causality attribution. Typecheck/lint/docs/budgets and 35 latest-base CPU contracts pass. Full root/all-template/startup/native qualification remains open. See [merged-source acceptance and retained report seals](../../verification/prd345/pr440-acceptance-2026-10-05.md). No further timing is admitted until a comparable window is available.
+
 ### Phase 3 — clean starter and platform validation
 
 - [ ] Verify generated clean starters/simple scenes and template visual gates including dark environment/no-sun and backlit character controls. proof: `pnpm test:templates`, tone crops and red-green captures
 - [ ] Complete required root and native qualification without claiming unexecuted targets. proof: `pnpm typecheck && pnpm lint && pnpm test` and targeted native fixture
+  - 2026-10-04 interim, open: typecheck and lint pass. `pnpm test` stops in runtime-native (21 reds, native host binaries not built here). Root vitest 8007 pass, 4 red under load 35-50: template typecheck passes in isolation, `generated-shooter-input.spec.ts` red twice and not yet attributed.
 
 ## Qualification design
 
@@ -331,7 +337,7 @@ source-first architecture requires an explicit coherent policy change before pac
    via `assert.tone`, that `p1` and `p99` remain separated — a scene lit only by the fill still has
    range.
    *Red-green:* set the analytic fill to black; the range assertion goes red.
-- [ ] **Every template has both, and says so.** `scripts/__tests__/primary-docs.spec.ts` and the proof: `primary-docs and generated convention rows`
+- [x] **Every template has both, and says so.** `scripts/__tests__/primary-docs.spec.ts` and the proof: `primary-docs and generated convention rows; 2026-10-04 primary-docs.spec.ts + instruction-budget.spec.ts 17/17 pass, spec asserts rimGain and fillGain in all 13 template AGENTS.md (rain included, custom shader exception stated)`
    templates' own gate assert each template's `AGENTS.md` names both conventions and both override
    names.
 - [ ] **The templates gate is green on the templates that currently pass it.** Note the known red lane: proof: `complete template gate with explicit lane outcomes`

docs/benchmark/prd345/shooter-causes1/README.md

@@ -0,0 +1,11 @@
+Shooter causal diagnostics, frozen before production integration
+
+The exact 1920×1080 performance scenario retained its 33 ms limit. Three balanced rounds bracketed five controls with the current source, producing 21 fresh launches: six current controls and three launches per other arm. All raw reports are retained as deterministic gzip files; manifest.json records original report and all 42 original PNG hashes. Twelve representative full PNGs are copied unchanged. contact.png scales selected originals to 50% for inspection. No original evidence was deleted.
+
+Mean frame-time p95: current 46.93 ms (0/6 pass), runtime-zero 51.63 ms (0/3), copy-only 29.37 ms (3/3), material-emissive-only 28.90 ms (3/3), compiled-rim-only 45.87 ms (0/3), shared full graph 30.43 ms (3/3). summary.json records every result, sample count and variance; the 62.7 ms runtime-zero sample is retained. Lower-cost diagnostic arms omit terms and are not default qualification.
+
+The fixed outer shared lease used /tmp/threenative-playtest-capture, holder PID 2521870. Parent TMPDIR was unset; child temporary data stayed on durable storage with its normal private lease. No own builds ran during timing. Noncooperating external host activity remains uncontrolled. This comparison used the same frozen installed runner for all arms; the later root runner signal repair has a different identity.
+
+Independent inspection of all three contact pairs and the first full images found consistent facade palette, window and doorway materials, gun/body shading and shadows. HUD/minimap presentation varies. Camera helper source is byte-identical, but the unchanged performance scenario requests player observations and does not establish exact runtime camera-matrix equality. These images support a coarse appearance check, not complete material/lifecycle qualification.
+
+Shared graph clears the unchanged cost limit in all three launches; fixed-pose causal negatives, multi-material fidelity, live-key/multi-camera behavior, disposal/recycle and repeated startup p95 still require postintegration qualification. The isolated experimental native lane does not admit production native defaults.