fix(engine): SDR content in HLG exports sits at reference white, not peak - #5749

Review at 790a486f.

Strengths

• alphaBlit.ts:308: the fix derives the scale from the HLG inverse OETF at 75% signal rather than hard-coding a constant. The scene light it gives is 0.26496, which is 203.15 nits on a 1000-nit display through the γ1.2 OOTF, matching BT.2408. HLG now follows the same reference-white model as the PQ branch's 203 nits.
• The Python generator mirrors the change exactly. I ran generate-lut-reference.py --probes from this head and its output is byte-identical to the regenerated SRGB_TO_HDR_REFERENCE (sRGB 255 → 49151 = 0.75 × 65535, 128 → 27146).

Verification

• Math: sRGB 18% grey lands at 37.8% HLG signal, which matches BT.2408's 38% grey. Reference white at 75% is 721 as a 10-bit limited-range code, consistent with the measured 724 in the body.
• PQ and SDR paths unchanged: only the transfer === "hlg" branch of buildSrgbToSignalLut changed. Every PQ probe value in the table is identical, the srgb branch is untouched, and no encoder or service file is in the diff.
• Colour metadata: unchanged and correct for HLG. chunkEncoder.ts:436-446 and streamingEncoder.ts:404-412 write bt2020nc / bt2020 / arib-std-b67 / tv.
• Sibling sites: the other HLG OETF, in shaderTransitions.ts:47, round-trips pixels that are already HLG signal (signal → linear → signal). It never takes SDR-relative light, so it doesn't need the scale.
• SDR clips: these are never in nativeHdrVideoIds (extractVideosStage.ts:437). They reach the canvas through the DOM capture and blitRgba8OverRgb48le, so the fix covers them too.
• Tests: the byte-exact probe table plus the white, 50%-alpha, near-opaque and asymmetric-channel tests all pin 49151. The BT.2408 invariant test asserts both transfers. The goldens are unchanged, and the body explains why: the HLG fixture renders no SDR content.

Notes (non-blocking)

1. A graded HLG clip still goes HDR → SDR (tone map) → grade → SDR → HLG, and those two conversions aren't inverses. The body's table shows warm-daylight moving the 75% patch from 720 to 627, so graded HLG clips now read darker than ungraded ones, where before they blew out. That's much better than before. A follow-up test pinning that an identity grade keeps an HLG clip within a few codes of its source would stop this drifting again.
2. alphaBlit.ts:308 and generate-lut-reference.py:26 each compute the constant. The drift rule in the generator docstring covers that, so this is fine as it is.

CI at 790a486f: all 8 required checks are green (Build, Render on windows-latest, Semantic PR title, Test, Test: runtime contract, Tests on windows-latest, Typecheck, regression).

Verdict: APPROVE
Reasoning: The HLG mapping now matches BT.2408 reference white, which I checked numerically and confirmed with a byte-identical regenerated table. The PQ and SDR paths and the HLG colour tags are untouched, and required CI is green at this head.

— Via