feat(templates): qualify bounded backlight and dark-environment defaults

[jonit-dev]

Generated subjects retain a restrained key-colored edge under backlighting, and dark scene environments receive a measured directional fill. Defaults now live in twelve actual template source trees, with live material enrollment and owned cleanup. Bright photographed environments suppress the extra fill; template-specific overrides preserve authored lights and palettes.

Tracker G09 (row 13), PRD-345. Hardware desktop WebGPU high tier qualifies; native, mobile, software and WebGL retain original materials. Custom materials stay original. This is source-first template behavior; automatic bare-engine material treatment is not implemented.

Real before / after

All twelve full pairs use matched source scenes, fixed tick-60 cameras/state, cooked assets, 1280×720, 4× MSAA and NVIDIA Turing hardware WebGPU. Root and an independent reviewer inspected all 24 frames. The default openings are deliberately subtle; the controlled backlight/black-environment fixtures below isolate the visible improvement.

Racing preserves the original dark road/shadows after fixing shared light-matrix writes during rendering. Puzzle keeps its original authored lighting with added gains zero; sailing's bright sky suppresses added fill. All twelve full pairs and provenance. Presentation-time flag/smoke phases and a small tower HUD hover variation are disclosed, not claimed as lighting improvements.

Controlled effect	Disabled	Enabled
Backlit edge
Dark environment, no sun

Six intended red/green fixture arms qualify: edge p99 53 versus 7 with rim disabled; body 2 versus 0. Black-fill and omitted-marker controls fail their exact intended assertions. Failed software-harness and rejected racing attempts remain retained.

Verification and measured costs

• Affected CPU suite: 65 files / 1011 tests pass; root typecheck and lint pass. Dynamic/lifecycle/read-only matrix tests pass, including restoration, shared ownership, recycling and tier changes.
• Current runtime source: 80/80 fresh Chromium launches, 40 per arm in ABBA order. Ready p95 2649.7→2831.0 ms (+181.3 ms, 6.84%); median 2560.3→2760.0 ms. Fresh process/context/page/HTTP cache; OS and GPU driver caches uncontrolled. No shader-cold or compile-complete claim.
• Actual GPU timestamp windows: dark 11.0467→11.7467 ms (+0.700 ms, 6.34%). Backlit negative delta is noise, not a speedup. Only 23 asynchronous samples per arm across three 60-frame windows, one serial run.
• Owned native host and V8/QuickJS prerequisites build; actual native unit tests pass 1531 with 62 skipped. Matched native opening images preserve original appearance: candidate lighting disabled, zero conversions, unknown radiance. Native antialias stage dropped (build:Image is not defined); network diagnostics waived because native has no observer. This does not qualify enabled native appearance.
• All thirteen clean generated projects boot. Original gameplay suites: 11 pass, shooter fails p95 52 ms > unchanged 33 ms (352 samples), snow fails 34 ms > 33 ms (223 samples). Quiet shooter controls complete: six original runs pass (per-run p95 mean 29.23 ms, range 28.1–30.2), six current runs fail (mean 48.17 ms, range 44.7–51.8). This is a whole-candidate regression under the matched setup; individual cause remains unproven. Snow originals pass 1/6 (mean 33.83 ms), current pass 5/6 (mean 29.72 ms); current is not reliably under the threshold. Historical durable-TMPDIR runs held private namespace leases with manual own-job serialization; shared global GPU exclusion was not established. Future timing uses an explicit outer shared lease.
• Normal root suite reached package tests but failed two packed-consumer fixture path checks under symlinked durable TMPDIR; canonical-root correction retains all identity assertions and all seven consumer tests pass after independent review. Full root rerun remains pending; unit phase was not reached. A second normal root attempt exposed an owned capture lease leak on SIGTERM; the narrow local signal repair passes 226 CPU tests and real forced/unforced cleanup gates. A separate pre-existing borrowed remote-page cancellation edge remains open.
• Evidence caps and 2574 relative documentation links pass. Source hashes, full raw reports, startup samples and failed attempts are retained in the PRD.

Remaining acceptance

Draft remains partial. Shooter has a reproduced matched performance regression; snow performance and the full root rerun remain open; all failures are preserved. Historical GPU isolation limitations remain disclosed. The native fallback comparison is complete; original native gameplay visibility observations are unsupported and remain blocked. A separate compatible opening witness does not replace those gates. Rain's custom raymarched shader has authored hemisphere/flash lighting but lacks the qualified named conventions; its cloud reflection image is not a spherical environment measurement. Rain and enabled native appearance remain unqualified. No original acceptance criterion was relaxed. Draft CI skips are not green checks; no ready/100/auto-merge. Shared develop integration remains held for coordinated #397/#402 work; workbook statuses unchanged.

Native fallback comparison

Matched tick 86, camera/projection, 1280×720 and NVIDIA RTX 2080 native WebGPU. Full raw starter reports, failure counts and native limitations.