Perceptual quality controller - #2
Merged
Merged
Conversation
codegraph serve resolves the workspace from its own cwd; the explicit
--path ${workspaceFolder} argument was passed through unexpanded by some
launchers and broke indexing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
The outside advisor's answers to the one-shot brief (verified findings, PR 0-5 program, gates). Registered via AKR source tooling; the root-level working copy was byte-identical and is removed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
…--quality seed The one-shot program's policy-side PRs, per the registered 2026-08-24 advisor memo: - PR 1: jpxl.quality-trace/2 groundwork — QualityWork counters (pixel plans, reconstructions, metric evaluations, entropy trainings, emissions) threaded through probe/price/rescue/reducer and folded across trials; prediction and transform_features blocks on QualityOutcome. - PR 2: sweep_frame_perceptual — fresh production-policy pixel plans at pinned effective-scale rungs for oracle labelling; effective_scale / rung_for_effective_scale made pub. - PR 4: TransformFeatureSummary — ten quantizer-independent DCT8 features reduced per LF group in parallel through the shared candidate forward cache (deterministic across worker counts; in-search the prefill warms the cover's own cache and measured net-neutral wall). - PR 5: the generated crossing model (qpv2-st-1, source + transform features, 130 image families) seeds the navigator — risk-adjusted candidate when confident, median when uncertain, legacy table only out of distribution — and its local slope prior steers the first correction under unchanged probe/price caps and canonical verification. one-shot-controller joins the crate's default features after the promotion A/B on 441 never-tuned holdout cells: zero floor violations on either arm, byte geomean 0.997 (locked holdout) / 0.990 (painting families), reconstructions -18% and wall -16% on the paintings, 12 MP anchor wall -21%. Also closes the PR 7 reducer question's stale docs and corrects the probe-cap accounting docs (observable maximum is pixel_probes + 1 including the rescue probe). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
…quality-ladder PR 0 of the one-shot program: an encode whose bounded search cannot verify the requested score now fails at the public surface — Error::TargetNotMet carrying a QualityMiss (kind, requested and best verified scores, probe/price counts, metric version, trace), CLI exit 1 with no output file — instead of writing an under-target stream with exit 0. Encoder::with_quality_fallback / --quality-fallback opt into the two explicit alternatives: lossless (mathematically lossless stream, status fallback_lossless) and best-effort (the previous behavior, now explicit, under its true status). A refused encode still appends its quality-trace record so failed searches remain calibration input; every met-path stream is byte-identical. Calibration surfaces for PR 1-4: the trace is now jpxl.quality-trace/2 (work, prediction and transform_features blocks; /1 stays readable by the harness); new 'jpxl quality-ladder' subcommand (fresh production-policy plans at pinned effective scales, canonically scored, optionally exact-priced, JSONL); 'jpxl features --transform-summary'. The CLI forwards the policy crate's one-shot-controller feature. Decision record: jpegxl-rs.decision.quality-miss-fallback-semantics. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
…ake, promotion A/B The one-shot program's PR 2-3 calibration pipeline and the promotion instrument: - quality_oracle_labels.py: resumable ladder sweeps (sweep) and censoring-aware label extraction (labels) — coarsest-meeting-rung, log-loss interpolated crossing, local slope, priced neighbors; --time-budget-minutes gives a clean resumable stop for time-capped corpus runs. - quality_predictor_v2.py: the pooled ln-loss crossing trainer — IRLS smoothed-pinball quantiles, robust standardization, ridge by held-out p90, family-grouped CV (LOFO or salted 10-fold), a blind never-tuned ext-holdout split excluded from tuning and from the shipped model, route simulation, and Rust codegen for the policy crate's generated predictor (run cargo fmt after regenerating). - quality_corpus_extend.py: external-corpus intake — family grouping from filename stems, deterministic sha256 cal/dev/ext-holdout split, round-robin per-directory sampling, PPM conversion, a manifest under jpxl.codec-corpus-ext/1 (originals are never modified). - one_shot_promotion_ab.py: interleaved A/B of two jpxl builds over manifest splits — floor violations, byte geomean, work counters, wall — the memo section 11 promotion screen. - make-quality-guard-fixtures.py: family identity fields (family_id/variant_id/generator_family/source_capture_id), a guard that fails the build if any family crosses a split, and nine new fixtures (56 images / 38 families). - codec_compare.py: reads quality-trace/2 (and still /1), passes --quality-fallback best-effort so RD sweeps grade under-target cells instead of dying on the new hard floor. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
Ledger for the memo-driven session: the quality-miss fallback-semantics decision; evidence for the falsification gate (honest FAIL on the 38-family corpus: LOFO median .22/p90 .81), the expanded-corpus gate (blind painting holdout median .113/p90 .258, PASS) and the promotion A/B (441 never-tuned cells, zero floor violations, PASS); the pqc-one-shot-controller work record at rev 2 with the open items (12 MP wall anchor 6.31x vs the 2.0x requirement, saturated/banding corpus coverage, 108 unswept ext-polart images); the proposed JPEG bitstream-recompression work record; and the regenerated views. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
Five gated phases from ISO/IEC 18181-2 s9.11 + Annex A and 18181-1: a new clean-room jpxl-jpeg coefficient codec (self-gated byte-exact JPEG round-trip), YCbCr 8x8-only VarDCT coefficient carriage with RAW dequant matrices, the jbrd box + Annex A reconstructor, CLI/facade surface, and black-box cjxl/djxl interop. Records the verified encoder prerequisites (no do_YCbCr / jpeg_upsampling / RAW-quant writing today) and the open questions (LF/DC bijectivity proof is the highest risk; 10918-1 source acquisition; OCR-suspect jbrd tables to check against the original scan; Brotli dependency decision). The work record jpegxl-rs.work.jpeg-bitstream-recompression now carries the plan as origin source and the four phase gates as acceptance checks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
… on large frames The 12 MP quality-mode wall was dominated by two serial per-probe stages inside render_metric: the varblock reconstruction (a 12 MP frame is a single LF group, so group-level parallelism gave it nothing) and the transfer-curve linearization. Both are independent per-element maps, so both now run banded over the EncodeExecutor with byte-for-byte identical output at any worker count: - varblocks: per-varblock compute (I.5.3/I.6/I.8/I.9 + J.4.3 sigma) moved into map_ordered chunks of 2048; each chunk's samples and sigma writes scatter serially into disjoint frame regions, and the dequant-matrix cache is warmed up front so the closure reads it through a shared borrow. Chunking bounds transient memory to a few MB; the locked 12 MP peak-memory anchor is unchanged (1,899 MB). - linearization: into_linear_rgb_at_depth_with bands the per-sample LUT lookup; the large-frame scoring path in jpxl-perceptual uses it. Verified byte-identical streams before/after on six image/target cells, identical canonical scores, cross-thread determinism (threads 1/4/8, fast-debug and release), and green fmt/clippy/tests on the touched crates. Release wall on the 12 MP anchor at 8 threads: 4.92 s -> 3.91 s (-21 %); matched-rate ratio 3.11x -> 2.46x on the measuring host. The remaining per-probe residual is split between the varblock scatter and the already-parallel metric. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
New observation pqc-large-frame-render-parallel-2026-08-25 (phase attribution, exact wins, -21% release wall, ratio 3.11x -> 2.46x on the measuring host, unchanged peak memory) and pqc-usable-efforts-cost rev 6: the wall check remains the sole open item, with the ranked residual directions — exact scatter overlap, blur fusion (risky), and the score-changing reduced-pyramid navigation that would need the Contract B promotion screen. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
The jpxl-jpeg crate glob matched no tracked path (V-102) — the crate does not exist yet; the plan document carries the intended layout. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Tct9rZempR3FTEksR9t3VX
The per-chunk varblock results were scattered into the frame serially after the parallel per-varblock compute. Each chunk's samples are now bucketed by the 64-row bands its varblocks intersect (at most two per varblock) and written by the executor's band workers; each sample belongs to exactly one varblock and one band, so the placed pixels are byte-for-byte the serial render at any worker count. Sigma writes are a few hundredths of the sample volume and stay serial. Measured honestly: wall-neutral at the 12 MP anchor (render_metric 2043/2168 ms before vs 2107/2057 ms after at 8 threads; 4 MP varblock stage 45 ms at 4 threads in both trees) — the ~50-100 ms/probe scatter estimate in the work note was high; the serial residual was ~30-45 ms and inside run noise. Kept because it removes the last serial per-sample stage, which grows with frame area (~150 ms/probe at 50 MP), at no measured cost. Byte-identity verified against the pre-change binary on six image/target cells and across threads 1/4/8 and AVX2 on/off; the full 52-cell locked production-identity matrix passes (52/52 identical). AKR-Change: chg-fff37cc35072bbe3 AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Graph: sha256:55e79d09551a4c24e5d889fe846ff51f9fe0154a11423475bc7554fc025ffcfd AKR-Tree: d43a520 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012PGFGWnmbi5WRCwCLu7s6K
Closes the one-shot memo's section-10 generation gap for the two critical classes where worst-cell byte regressions concentrated: saturated (4 -> 26 families) and gradient/banding-stress (8 -> 25). Every new family is a distinct deterministic (seeded) content process — hue wheels, angled saturated bands, near-clipping chroma ramps, complementary checkerboards, saturated noise, sub-kilobyte-encodable solid tiles, concentric rings, out-of-gamut clip plateaus, Voronoi; shallow luma/chroma ramps, dithered ramps, radial sky glows, sunset stops, bilinear fields, conic sweeps, diamond gradients — not one recipe with 25 seeds. Corpus grows 56 -> 95 fixtures (calibration 43, development 32, holdout 20); the split audit shows no family in two splits, regeneration is sha256-idempotent, the manifest validates, and all 13 previously locked holdout fixtures are byte-unchanged, so existing evidence stays comparable. 16 generator unit tests pass. Label sweeps over the new families are deliberately not run here. AKR-Change: chg-c88603675047fe63 AKR-Work: jpegxl-rs.work.pqc-one-shot-controller AKR-Graph: sha256:55e79d09551a4c24e5d889fe846ff51f9fe0154a11423475bc7554fc025ffcfd AKR-Tree: 92aa67c Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012PGFGWnmbi5WRCwCLu7s6K
- README: verified against CONFORMANCE and the CLI; the quality-search section now describes the promoted one-shot seed (qpv2-st-1 pooled quantile regression behind the default `one-shot-controller` feature, with the calibrated-table fallback for out-of-envelope inputs) and the full-resolution canonical verification before every emission. - JPXL/docs/LICENSING-AUDIT.md: MIT-only audit of all 12 workspace crates and the full dependency graph; no copyleft anywhere; the only non-MIT arms are permissive (moxcms/pxfm BSD-3/Apache-2.0 on the default path via the image adapter; BSD/CC0 metric crates behind off-by-default measurement features). - JPXL/tools/bench_vs_libjxl.sh: dependency-light reproducible size/SSIMULACRA2/wall comparison (jpxl vs cjxl/djxl) with provenance headers (versions, sha256, flags), JSONL output, and clearly marked JPXL-only degradation when no runnable native oracle is present. Smoke-tested; native Linux cjxl/djxl (setup-oracles.sh) is still needed to exercise the comparison columns on this host. AKR-Change: chg-3c1cf97d55c892ef AKR-Work: jpegxl-rs.work.publish-readme-benchmark-mit AKR-Graph: sha256:55e79d09551a4c24e5d889fe846ff51f9fe0154a11423475bc7554fc025ffcfd AKR-Tree: cdc6873 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012PGFGWnmbi5WRCwCLu7s6K
New clean-room crate implementing ISO/IEC 10918-1 at the coefficient level: markers, quant and Huffman tables, frame/scan headers, and entropy-decoded DCT coefficients, re-emitted byte-for-byte. Baseline (SOF0), extended sequential (SOF1), and progressive (SOF2, all four procedures) are covered; restart intervals capture per-segment padding; APPn/COM/unknown markers and post-EOI tail data are preserved verbatim; arithmetic, hierarchical, lossless, and 12-bit inputs are refused with a typed error. No workspace or third-party dependencies, no unsafe, hand-rolled error enum, unit-bearing newtypes. The one non-derivable wire fact is recorded rather than re-derived: libjpeg splits long progressive AC end-of-band runs when its correction-bit buffer fills, and that split is invisible in the coefficients, so decode records the exact EOB-run lengths per scan and encode replays them flush-on-match. A regression fixture with split runs proves the recorded runs are load-bearing. Verified over the full private Pol Art archive via the deterministic roundtrip_sweep example: 5,217 of 5,218 JPEGs byte-identical, zero mismatches (coverage: 3,166 baseline, 2,049 progressive, 2 extended sequential, 4:4:4/4:2:0/4:2:2/4:4:0, grayscale, 684 restart-interval, Exif/ICC/XMP, 41 with trailing data); the one remaining file is a truncated zero-padded JPEG that libjpeg's djpeg also rejects, refused here with a typed error. 34 crate tests pass (17 bit-exact roundtrips on generated fixtures with provenance sidecars, 10 typed refusals, 2 EOB-run regressions, 4 unit, 1 doctest); workspace gates green. AKR-Change: chg-2089f6782d8c12a3 AKR-Work: jpegxl-rs.work.jpeg-bitstream-recompression AKR-Graph: sha256:6eb74db06baef609d12d358f1a48e5fe3a85a55f9db79ea04ef192b562d021f7 AKR-Tree: 6b253d1 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012PGFGWnmbi5WRCwCLu7s6K
- pqc-usable-efforts-cost rev 7: memory-12mp (1.95 GiB peak) and production-identity (52/52 cells, threads 1/4, AVX2 on/off) re-verified on the current binary; the quiet-host wall observation re-baselines the anchors (12 MP 5.42x/4t, 4.25x/8t) and shows the 2.0x wall target is unreachable by search improvements alone - fixed costs are ~1.95x the rate path; wall-anchors stays open pending a target decision. - pqc-one-shot-controller rev 3: the saturated/gradient corpus generation gap is closed (26/25 families); label sweeps remain. - jpeg-bitstream-recompression active: Phase A evidence over the full 5,218-file archive; the EOB-run record/replay discovery is noted for Phases B/C. - publish-readme-benchmark-mit completed: all four checks evidenced (reproducible native-oracle comparison, MIT-only audit, README current-state, workspace gates). - papercut: akr commit-msg hook rejects message-only amends after the change transaction closes. AKR-Change: chg-186f2acbbc075c3f AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Evidence: @jpegxl-rs.evidence.bench-vs-libjxl-reproducible-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.jpeg-phase-a-roundtrip-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.mit-only-licensing-audit-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-memory-12mp-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-scatter-production-identity-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.readme-current-state-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.workspace-gates-2026-08-25/1 AKR-Graph: sha256:30546cc5534a77086da577a50522e7178dea87561649cc48bd29bbec56746c56 AKR-Tree: 3511116 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012PGFGWnmbi5WRCwCLu7s6K
Profile and reduce the production Fast and Balanced perceptual path's full-frame render/metric allocation and rescue-probe cost. Work only on usable efforts: do not spend measurement time on the feature-gated Quality reference effort. Pure scorer, renderer, and lifetime changes must preserve Fast/Balanced codestream bytes; any deliberate search-policy change requires the standing Contract B screen. AKR-Change: chg-9465cc41a44894e8 AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Graph: sha256:30546cc5534a77086da577a50522e7178dea87561649cc48bd29bbec56746c56 AKR-Tree: ed9e0af
- jpegxl-rs.papercut.a-work-revision-that-cites-already-committed new -> verified No AKR work record: ledger maintenance: papercut record and regenerated views only AKR-Change: chg-d73accb685d2f11e AKR-Graph: sha256:ef95ff5110be9aeb247bd9805fd425acc01148ef91bb9ca4e8c9dd0c7efb3f8c AKR-Tree: 7fceaad
…paths Four value-preserving optimizations in the SSIMULACRA2 probe path, found by phase attribution at the 12 MP anchor (per probe at 8 threads: render 195 ms, score 290 ms, of which scale-0 candidate blurs ~160 ms): - The vertical blur pass writes each column strip's disjoint columns of the output in place (DisjointColumns), replacing per-strip buffers, a mutex collection, and a serial full-plane scatter copy. Per scale-0 blur the vertical half dropped from ~10 ms to ~4.7 ms. - The horizontal blur pass runs four rows at a time as independent recursion lanes over a lane-interleaved padded row, the row analogue of the vertical pass's column lanes, dispatched to an AVX2 build like the vertical pass. Per lane the arithmetic is exactly the scalar step's, in the same order. - Redundant re-zeroing is dropped where every element is provably overwritten before it is read: the blur's horizontal temp (re-memset up to nine times 48 MB per 12 MP probe) and the mu2/s22/s12 and recomputed-reference scratch planes. - from_srgb16 evaluates the transfer function through a 65536-entry table of the same expression instead of a powf per sample, and deinterleave preallocates its planes. Exactness: 6/6 A/B cells byte-identical vs the pre-change binary (threads 1/4/8, AVX2 on/off), full locked production-identity matrix 52/52 identical, 12 MP output bytes unchanged. Measured on the standing interleaved recipe (load_1m 1.8-4.5): quality/rate 3.40x/2.76x at 4.3 MP (4t/8t) and 5.07x/3.50x at 12 MP, from 3.87x/3.22x and 5.42x/4.25x; 12 MP peak RSS 2,008,188 KiB from 2,043,056 (ceiling 2,097,152). Artifacts: .agent/scratch/pqc-metric-cuts-20260825 (kept). AKR-Change: chg-24eecb5fd4bda387 AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Graph: sha256:ef95ff5110be9aeb247bd9805fd425acc01148ef91bb9ca4e8c9dd0c7efb3f8c AKR-Tree: a9240f5 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012PGFGWnmbi5WRCwCLu7s6K
Profile and reduce the production Fast and Balanced perceptual path's full-frame render/metric allocation and rescue-probe cost. Work only on usable efforts: do not spend measurement time on the feature-gated Quality reference effort. Pure scorer, renderer, and lifetime changes must preserve Fast/Balanced codestream bytes; any deliberate search-policy change requires the standing Contract B screen. - jpegxl-rs.evidence.pqc-metric-cuts-memory-12mp-2026-08-25 new -> verified - jpegxl-rs.evidence.pqc-metric-cuts-production-identity-2026-08-25 new -> verified - jpegxl-rs.evidence.pqc-metric-cuts-wall-anchors-2026-08-25 new -> verified - jpegxl-rs.evidence.pqc-metric-cuts-workspace-gates-2026-08-25 new -> verified Verified by: - Metric-cuts 12 MP memory: peak 2,008,188 KiB under the 2 GiB ceiling - Metric-cuts production identity: 52/52 matrix and 6/6 pre-change A/B byte-identical - Metric-cuts wall anchors: ratios cut ~9-19 percent, 2.0x target still unmet - Metric-cuts workspace gates: fmt, clippy -D warnings, release tests all pass AKR-Change: chg-ae2be006272a0327 AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Evidence: @jpegxl-rs.evidence.pqc-metric-cuts-memory-12mp-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-metric-cuts-production-identity-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-metric-cuts-wall-anchors-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-metric-cuts-workspace-gates-2026-08-25/1 AKR-Graph: sha256:11b87518a2e357f90ab0354254cb1ca3fef23f1be56d7cd245257f95d636a6a5 AKR-Tree: 7036c46
- jpegxl-rs.papercut.knowledge-get-with-detail-canonical-on-jpegxl new -> verified No AKR work record: papercut-only ledger change names no work record AKR-Change: chg-cc21e841d39f4cd6 AKR-Graph: sha256:c5469183a640fdfb5dd28a9a6034632a0c2f9e378d5469f0dd6a6c1dd597d679 AKR-Tree: a9df958
…pth-quantized probe render Guided by a real cycle profile (perf, enabled by the operator): after the blur cuts, `__powf_fma` was 5.4% of cycles — a per-sample sRGB `powf` encode in the render's colour pass whose result every probe immediately quantized away, plus a per-sample transfer `powf` in the 16-bit prep — and the two-step linearize pass another ~3%. - PlanRenderer::render_linear_at_depth_with renders straight to the depth-quantized linear planes a probe scores: each linear sample's integer level is found by comparing against thresholds bisected over the f32 bit lattice from the actual linear_to_srgb -> scale -> round -> clamp composition, then mapped through the existing linear table. A guide table over the bit lattice pins each sample's level to a bucket-local range (two loads and at most a short ordered scan), and the classifier is cached on the renderer per bit depth so thresholds are bisected once per search, not per probe. The signalled sRGB encoding is never materialised on this path; render_with and RenderedFrame are unchanged for every other caller. - The classifier's equivalence with encode-then-round-trip is pinned by fused_depth_levels_match_the_srgb_round_trip: a dense sweep, every threshold's bit-lattice neighbourhood, both transfer branch seams, and the non-finite specials (which take level 0 exactly as the round trip's finiteness guard does). - PlanRenderEvaluator scores through the fused path on both the standard and the large-image branches. - PreparedFrame::from_srgb16_with evaluates its transfer function through a 65536-entry table of the same per-value expression. Exactness: 6/6 A/B cells byte-identical vs the pre-change 1fc3ca7 binary (threads 1/4/8, AVX2 on/off), locked production-identity matrix 52/52 identical, 12 MP bytes unchanged. Quiet-host 12 MP 8-thread quality wall 2.39-2.53 s (2.62 s before the fusion); interleaved-recipe ratios 3.06x/2.50x at 4.3 MP (4t/8t) and 4.65x at 12 MP/4t, from 3.40x/2.76x and 5.07x this morning (the 12 MP/8t schedule cell was load-polluted; see the kept scratch README). Peak 12 MP RSS unchanged. Artifacts: .agent/scratch/pqc-metric-cuts-20260825 (round 2). AKR-Change: chg-950455e93e20d5ce AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Graph: sha256:c5469183a640fdfb5dd28a9a6034632a0c2f9e378d5469f0dd6a6c1dd597d679 AKR-Tree: c6a9fe7 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012PGFGWnmbi5WRCwCLu7s6K
Profile and reduce the production Fast and Balanced perceptual path's full-frame render/metric allocation and rescue-probe cost. Work only on usable efforts: do not spend measurement time on the feature-gated Quality reference effort. Pure scorer, renderer, and lifetime changes must preserve Fast/Balanced codestream bytes; any deliberate search-policy change requires the standing Contract B screen. - jpegxl-rs.evidence.pqc-fused-render-memory-12mp-2026-08-25 new -> verified - jpegxl-rs.evidence.pqc-fused-render-production-identity-2026-08-25 new -> verified - jpegxl-rs.evidence.pqc-fused-render-wall-anchors-2026-08-25 new -> verified - jpegxl-rs.evidence.pqc-fused-render-workspace-gates-2026-08-25 new -> verified Verified by: - Fused-render 12 MP memory: peak 2,008,408 KiB under the 2 GiB ceiling - Fused-render production identity: 52/52 matrix and 6/6 pre-change A/B byte-identical - Fused-render wall anchors: 4.3 MP at 2.50x (8t), 2.0x target still unmet - Fused-render workspace gates: fmt, clippy -D warnings, release tests all pass AKR-Change: chg-65a35814006d248f AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Evidence: @jpegxl-rs.evidence.pqc-fused-render-memory-12mp-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-fused-render-production-identity-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-fused-render-wall-anchors-2026-08-25/1 AKR-Evidence: @jpegxl-rs.evidence.pqc-fused-render-workspace-gates-2026-08-25/1 AKR-Graph: sha256:cd475cd59072daec056b7dd990d7a577cabe3bed751c45724d18b3e51111151d AKR-Tree: 0a22526
Add the two metadata seams the archive pipeline needs:
- append_exif writes a Part 2 Exif box (LBox + "Exif" + u32
tiff_header_offset + raw TIFF payload) after the codestream boxes;
the facade exposes Encoder::with_exif, which validates the TIFF
byte-order header and implies the container.
- ColourSpace {Srgb, LinearSrgb, DisplayP3, Rec2020} threads through
EncodeOptions into the Annex E ColourEncoding bundle (Table E.1
non-default sequence). Lossless Modular signals colour declaratively;
lossy VarDCT rejects non-sRGB with a typed Unsupported error because
the XYB transform and the metric are defined on sRGB input.
- Facade wrapping is unified in finish(); the rgb8/gray8 paths no
longer widen u8 samples through an intermediate Vec<u16>.
Round-trip tests cover the box walk, all four colour spaces through the
decoder, the greyscale guard, and TIFF-header rejection.
- jpegxl-rs.work.pipeline-metadata-seam new -> proposed
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QsAR4R1kGCRk8spmhQZMC2
AKR-Change: chg-2985272a860c003a
AKR-Work: jpegxl-rs.work.pipeline-metadata-seam
AKR-Graph: sha256:e56b9a85583615d94e1fab8b121fb543a4e24563fe57c169a07ccbb4dcc093fd
AKR-Tree: 7985656
accumulate_band now dispatches to an avx2+fma multiversioned clone when the CPU supports it. The lane loop computes the per-pixel SSIM' and asymmetry terms in POOL_LANES groups, then folds them into the f64 MapSums in original pixel order, so the accumulated sums -- and the encoded bytes -- are identical to the scalar walk (verified by test across remainder lengths and by cmp on full encodes). ~5-8% off the 12 MP quality-path wall. jpxl-plan-render now names the jpxl-core "simd" feature it depends on instead of relying on feature unification. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QsAR4R1kGCRk8spmhQZMC2 AKR-Change: chg-35a2d60333479cdd AKR-Work: jpegxl-rs.work.pqc-usable-efforts-cost AKR-Graph: sha256:e56b9a85583615d94e1fab8b121fb543a4e24563fe57c169a07ccbb4dcc093fd AKR-Tree: 825f2bb
tools/build-release-final.sh runs the measured PGO cycle (instrumented build, training encodes over the canonical test set across the quality, rate and lossless paths plus a decode, llvm-profdata merge, profile-use rebuild) on the new release-final profile. Measured today the combined fat-LTO+PGO gain over `release` is 2-8%; the ledger's 36-53% PGO evidence predates the fused-render and Opt-F..P rounds and needs re-verification. `release` (thin LTO) stays the benchmark profile. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QsAR4R1kGCRk8spmhQZMC2 AKR-Change: chg-105dcc145ce18786 AKR-Graph: sha256:e56b9a85583615d94e1fab8b121fb543a4e24563fe57c169a07ccbb4dcc093fd AKR-Tree: 58d9bd8
The metadata seam is verified end to end at this commit: an appended Exif box round-trips through the decoder's box walk with the codestream untouched; with_exif rejects payloads - jpegxl-rs.evidence.avx2-pooling-bit-identity-2026-08-26 new -> verified - jpegxl-rs.evidence.pipeline-metadata-seam-2026-08-26 new -> verified - jpegxl-rs.observation.pgo-gain-remeasured-2026-08-26 new -> verified Verified by: - jpegxl-rs.evidence.avx2-pooling-bit-identity-2026-08-26 - jpegxl-rs.evidence.pipeline-metadata-seam-2026-08-26 AKR-Change: chg-438b67ff073613c8 AKR-Work: jpegxl-rs.evidence.pipeline-metadata-seam-2026-08-26 AKR-Work: jpegxl-rs.evidence.avx2-pooling-bit-identity-2026-08-26 AKR-Work: jpegxl-rs.observation.pgo-gain-remeasured-2026-08-26 AKR-Evidence: @jpegxl-rs.evidence.avx2-pooling-bit-identity-2026-08-26/1 AKR-Evidence: @jpegxl-rs.evidence.pipeline-metadata-seam-2026-08-26/1 AKR-Graph: sha256:2e53454b4abc5a2e7803995f3fb00745b536d02f33e33c9607223cdf0dd3935b AKR-Tree: 3f08f2f Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QsAR4R1kGCRk8spmhQZMC2
Authoritative plan for jpegxl-rs.track.encoder-optimization: execute Opt-F through Opt-P in after-order, supported by the static advisor assessment and the optimization decisions and gates. AKR-Change: chg-d328f28ddbc84320 AKR-Work: jpegxl-rs.work.encoder-optimization-plan AKR-Graph: sha256:2e53454b4abc5a2e7803995f3fb00745b536d02f33e33c9607223cdf0dd3935b AKR-Tree: d7a6de0
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.