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

[jonit-dev]

The generated backlight shader now shares one render graph per controller across twelve material templates. Each converted material retains its own PBR inputs, maps and original emissive. A regression repair also resolves key-light layers against the actual render camera: when a second camera has different layers from the controller camera, it now receives the correct key contribution.

The controller owns the shared graph through a WeakMap. The render callbacks read current gains, light state and actual camera layers without changing light matrices. Rain keeps its custom shader.

The historical 21-launch study at unchanged 1920×1080 and 33 ms measured mean per-run p95 46.93 ms for per-material graphs (0/6 passes) and 30.43 ms for shared graphs (3/3). Raw reports, captured sources and PNG hashes remain in docs/benchmark/prd345/shooter-causes1/. This historical result does not admit the complete current source: runtime/package/cooked-model and resource differences remain, and the camera callbacks have since been repaired.

Current verification:

• Actual-camera regression was RED before repair; camera/matrix/lifecycle tests pass 31/31.
• Merged scaffold/compiler contracts pass 101/101, including all thirteen hashes measured through the real generator. Latest merged exposure/docs/instruction contracts pass 35/35.
• Ordinary JS/DTS build, root typecheck, lint, docs and budgets pass. Lint retains existing warnings; budgets reports existing native-census drift.
• Fresh local difficult-lighting PNGs pass the unchanged decoded-pixel qualifier and independent visual review: edge/body p99 52/2 versus zero-rim 7/0; fill p99 43 versus black 0. Black fails only the two intended tone assertions. The omitted-marker runner passes, and its external marker qualifier rejects the missing marker. Actual camera matrices match across all four assertion arms.
• Independent final source/evidence review found no remaining actionable source defect.
• Published head 30d7847bb passed CI37272442106. Its exact candidate 673639ae has the same tree. The complete run passed 65 jobs with one explicit Android payload publish skip. All thirteen template nonvisual/pristine-typecheck legs, all four root unit shards (8,155 passed / 40 skipped tests), package, native, browser, playtest, typecheck and lint pass. These CI gates do not replace hardware template appearance or the unchanged 33 ms scenario.
• Current Linux-x64 native starter and relocated/container captures each pass 300 frames. The native consumer passes movement, collection and restart. Source admission gates, CPU material-identity controls and actual enabled=false / convertedMaterials=0 execution jointly verify original-material preservation. Environment radiance honestly remains unknown and fill stays unadmitted. Independent review inspected the pixels and verified retained report/screenshot hashes. This is fallback execution and preservation evidence; matched current visual/cost observations remain unverified.

Current integrated performance acceptance is RED. Six unchanged 1920×1080 / warmup60 / 33 ms scenarios measured shooter p95 112.9 / 144.5 / 88.9 ms and snow 59.8 / 48.6 / 51.8 ms; one shooter run also failed minimum FPS. Draw and triangle limits passed. These runs used two physical helper cores on a busy host and do not isolate code causality. Every failed outcome is retained. A same-current-assets comparison on physical cores 6–8 and their siblings passed measured suitability checks, then an actual CI Runner.Worker arrived eight seconds into the per-material launch. That arm was aborted; the shared arm was unrun, and no completed report, screenshot or new p95 statistic was produced. A preceding preparation-only attempt was retained separately after an idle-listener poll exposed an overly strict harness check; the focused listener classification regression is now 14/14 green. The scenario thresholds stayed unchanged, and the capture stopped without an automatic retry.

Current acceptance checkpoint and report seals distinguish verified, failed and unrun criteria. PRD-345 phases 2–3 remain open: integrated gameplay/performance and repeated startup admission, the complete all-template hardware gate and matched visual/cost observations, multi-camera and mixed-material/disposal GPU pixels, and rain's custom convention. Native original-material preservation and targeted Linux-x64 execution are verified. Enabled native grazing/fill is outside the documented admission scope; it is not an extra completion requirement. No local native build or denied replay was attempted in this task.

Published head 30d7847bb preserves the owner reconciliation and develop d5d16970 by normal merges. Develop subsequently advanced to 9b3c4824 through animation PR #395; the recorded qualification belongs to the explicitly named source tree above, and no new-base qualification is claimed. All new images remain local pending sharing approval; none were uploaded or added to Git. No performance limit, assertion, deadline or hook was weakened or bypassed.