docs: render the accepted Character 1 idle at 492 pixels wide

CHECKPOINT.md

@@ -0,0 +1,38 @@
+# Examples #29 / PR #64 — final frozen revision
+
+State: PASS under HQ's owner-adopted measured-wrap bar; frozen for parent commit, P2 response, CI/review/sign. No active jobs; no further render or encoding after candidate measurement.
+
+## Canonical input
+
+Origin `40db33d91e4aacbe4cd31ac9f3cc557973b4a1d4`; accepted Character 1 `family/roa/out/check/build/`, hash-bound by `family/accepted-inputs.json`. Model SHA256 `b9b61d55614805e468b27f9fb2dbaba65a50a0ce56b32ffef04f8cacd2570632`. All original corpus files remain unchanged.
+
+Pins: rig-c 2.68.0, spine-core 4.3.13, Bun 1.4.2, FFmpeg 7.0.2-static. Source rig/model/motion/art untouched; no generation, hand pass or retarget.
+
+## Adopted result and native boundary
+
+**rigc emits no closure verdict; this is the measured wrap**
+
+Original max1290 producer logs have zero `loop:` lines. Renderer help supplies no closure flag; `core_commands.ts:1030–1092` prints sample count, physics and sidecar, and `physicsLine` at 1117–1130 counts constraints. Actual native output: `  ..    physics: 23 constraint(s) stepped at 12 fps`. Native closure is UNAVAILABLE, not PASS or FAIL.
+
+HQ adopted measured wrap MAE 0.44359887397323583 ≤ max adjacent 0.7597497951723703 as this card's bar: PASS. First/last max RGB difference 212 and 1,917 pixels over 40 are observations only; exact endpoint equality is NOT a bar. Repeat metadata is recorded separately.
+
+## Selected artifact
+
+- Predictor lossless APNG: 492×1290, 41 independently decoded frames, all RGBA bytes identical (max error 0), each delay exactly 1/12 s, infinite plays metadata 0.
+- 11,351,964 bytes, SHA256 `bd31ad5b7d952733417a8ac830c40e73719f95ef0cdb8569e092c2bde3161f0e`.
+- Original compressed reference: 17,453,151 bytes, SHA256 `462044902d2c1ac080d964238d3d05dab7eddafbe2c0940657cba8bf460c4ff7`, retained remotely throughout and after measurement. Reduction: 6,101,187 bytes, about 35%.
+- Palette APNG: 2,638,904 bytes, max RGBA error 66, timing exact; rejected.
+- GIF: 2,253,537 bytes, max RGBA error 66, 8/100 and 9/100-second delays; rejected for error and exact-timing failure. All candidates preserve 41 decoded frames and infinite metadata.
+- Full candidate hashes/delays/commands are in the trial receipt. No rerender for optimization.
+
+## Verification and P2
+
+- Optimization preflight `20261024T164354Z-ca8fa850a810460797563a2952311be5`: exit 0; actual source APNG decoder, palette readers, codec control and error controls 0/255.
+- Candidate measurement `20261024T164502Z-48352c7770aa49189a0445bc3f2d7819`: exit 0; selected measured artifact copied unchanged.
+- P2 discussion_r4367769592 at reviewed head b8422d447d7ae91586a61e849fc9b76d9dd1ee4f READ and fixed: absolute TRIAL from a trial-containing checkout, separate INPUT_ROOT with four pinned source hashes. Extracted README setup, fresh frozen install and actual renderer loader dry-run passed without rendering.
+- `encode.mjs compress` uses the exact measured predictor arguments. Final revision validation is limited to lightweight hashes, syntax, path/setup consistency and command-argument comparison; no new full encode run is claimed for this final wiring.
+- Prior 82 producer PNGs were hash-verified and deleted. Optimization used in-memory decoded RGBA; its temporary workspace was removed. Small logs and compressed candidates remain remotely. No original source deletion.
+
+## File ownership
+
+PR revision changes: CHECKPOINT.md plus trial README.md, encode.mjs, receipt.json and idle-2x.apng. The trial sheet.png remains unchanged. No family edits, git/public writes, new review requests, subagents or full suite. Parent owns commit, P2 response, CI, review and sign.

trials/2026-10-25-character-1-idle-2x/README.md

@@ -0,0 +1,209 @@
+# 2026-10-25 Character 1 idle at twice the recording scale
+
+## Question
+
+Does rendering the accepted Character 1 idle at a larger `rigc render --max`
+produce a lossless recording at least 492 pixels wide, retaining all 41
+12-fps samples and a measured loop?
+
+[APNG recording](idle-2x.apng) · [41-frame sheet](sheet.png) ·
+[Machine-readable receipt](receipt.json) · [Encoder and verification](encode.mjs)
+
+## Inputs
+
+The input is the current accepted office Character 1 build,
+[`family/roa/out/check/build/`](https://github.com/rig-c/examples/tree/40db33d91e4aacbe4cd31ac9f3cc557973b4a1d4/family/roa/out/check/build),
+at examples commit `40db33d91e4aacbe4cd31ac9f3cc557973b4a1d4`.
+The [accepted manifest](https://github.com/rig-c/examples/blob/40db33d91e4aacbe4cd31ac9f3cc557973b4a1d4/family/accepted-inputs.json)
+and the Family guide identify that serving build as canonical. The receipt
+records the exact model, skeleton, atlas and PNG hashes. They were read,
+not rebuilt or edited. The site's historical `c239b09` build is not silently
+substituted for this current accepted input.
+
+Withheld: nothing; the inputs already exist in the public corpus. New raw
+frame sets are reproducible outputs, deleted after the sheet verdict and
+concatenated-PNG hashes were recorded. Acquisition, pelvis, clean and retarget
+are not applicable: this trial reads an existing idle and acquires no video
+or new pose sequence.
+
+## Commands and pins
+
+| Tool | Pin |
+| --- | --- |
+| rig-c | 2.68.0 |
+| official spine-core | 4.3.13 |
+| Bun | 1.4.2 |
+| FFmpeg, independent decoder | 7.0.2-static |
+
+Run from a checkout that **contains this trial** (the PR or a later merged
+commit), not from the older input commit alone: that older commit predates
+`encode.mjs` and `receipt.json`. `TRIAL` is an absolute path to these new trial
+files. `INPUT_ROOT` may be this current checkout or a separate input checkout;
+the setup verifies all four accepted input hashes before any rendering. The
+historical commit identifies input provenance, not where the new script lives.
+
+Choose a new empty output path for `WORK`. Install only into that workspace.
+The exact package manifest and frozen Bun lock are preserved in `receipt.json`.
+
+```bash
+TRIAL="$PWD/trials/2026-10-25-character-1-idle-2x"
+INPUT_ROOT="$PWD"
+WORK="$PWD/idle-2x-reproduction"
+python3 - "$TRIAL" "$INPUT_ROOT" "$WORK" <<'PYSETUP'
+import hashlib, json, sys
+from pathlib import Path
+trial, input_root, work = [Path(p).resolve() for p in sys.argv[1:]]
+assert (trial / 'encode.mjs').is_file(), 'TRIAL must contain the new trial script'
+receipt = json.loads((trial / 'receipt.json').read_text())
+source = input_root / receipt['source']
+for name, expected in receipt['sourceInputHashes'].items():
+    actual = hashlib.sha256((source / name).read_bytes()).hexdigest()
+    assert actual == expected['sha256'], 'accepted input hash mismatch: ' + name
+work.mkdir(exist_ok=False)
+(work / 'tools').mkdir()
+(work / 'tools/package.json').write_text(json.dumps(receipt['toolchain']['packageJSON'], indent=2) + '\n')
+(work / 'tools/bun.lock').write_text(receipt['toolchain']['bunLock'])
+print('PASS: current trial files and all four pinned accepted input hashes')
+PYSETUP
+BUN_INSTALL_CACHE_DIR="$WORK/cache" bun install --cwd "$WORK/tools" --frozen-lockfile --ignore-scripts
+export BUN_INSTALL_CACHE_DIR="$WORK/cache"
+export BUN_RUNTIME_TRANSPILER_CACHE_PATH="$WORK/transpiler"
+bun --no-install "$WORK/tools/node_modules/rig-c/bin/rigc.cjs" render --candidate "$INPUT_ROOT/family/roa/out/check/build" --animation idle --fps 12 --max 1290 --poser core --out "$WORK/frames"
+bun "$TRIAL/encode.mjs" preflight "$WORK/frames" "$WORK/tools"
+bun "$TRIAL/encode.mjs" encode "$WORK/frames" "$WORK/tools" "$WORK/reference.apng"
+bun "$TRIAL/encode.mjs" compress "$WORK/reference.apng" "$WORK/tools" "$WORK/idle-2x.apng"
+```
+
+Heavy rendering and encoding were run on Nova. The staged order was render
+checkpoint → actual PNG/FFmpeg reader dry-run and metric controls → encode
+that same checkpoint → independently decode and compare all 41 frames.
+`rigc render --help` was read before choosing the flags. Inspect the sheet
+and retain the receipt before deleting the newly rendered frame directory.
+The script prints the concatenated PNG hash and a separately labelled measured
+seam. `compress` uses the exact FFmpeg predictor arguments measured for the
+selected candidate, then checks timing/repeat metadata and independent decoded
+RGBA equality. The original compressed reference was retained throughout
+candidate measurement. No rerender was used for optimization.
+
+The README setup was extracted and executed with absolute `TRIAL` and
+`INPUT_ROOT`: all four pinned source hashes, a fresh frozen install and the
+actual renderer model/skeleton/atlas/PNG loader passed. This setup dry-run did
+not render. It fixes PR #64 review
+[discussion_r4367769592](https://github.com/rig-c/examples/pull/64#discussion_r4367769592).
+
+## Measurement
+
+The bar was fixed before rendering: width at least 492, all 41 samples at
+12 fps, exact delay fractions and repeat metadata, and an explicit endpoint
+and wrap measurement. HQ adopted the measured-wrap acceptance criterion: wrap RGB mean absolute
+difference no larger than the largest adjacent RGB step. RGB ignores alpha;
+the renderer's background is opaque `(232,232,232,255)`. Equal frames are the
+zero control; a black-to-white pixel gives max/mean difference 255 and one
+pixel over 40. Zero motion is handled without dividing by a zero step.
+
+| Measured | Result | Bar / interpretation |
+| --- | --- | --- |
+| First producer, max 1280 | 488×1280 | FAIL width ≥492 |
+| Final producer, max 1290 | 492×1290 | PASS width ≥492 |
+| Renderer samples | 41 at 12 fps, t=0…40/12 s | All retained |
+| APNG decoded frames | 41, all RGBA pixels equal to source | PASS lossless |
+| APNG delays | 41 × exactly 1/12 s | PASS; no centisecond rounding |
+| Encoded duration | 41/12 s = 3.416666… s | Includes every sample, including endpoint |
+| APNG `acTL.num_plays` | 0 | Infinite repeat metadata only |
+| First vs last max RGB difference | 212; 1,917 pixels over 40 | Observation only; exact endpoint equality is **not a bar** |
+| Wrap mean absolute RGB difference | 0.44359887397323583 | ≤0.7597497951723703 maximum adjacent step: PASS smooth wrap |
+| Physics stepped during render | 23 constraints at 12 fps | Not a physics-disabled keyed-loop check |
+| Selected APNG size | 11,351,964 bytes | Predictor-optimized; lossless RGBA, error 0 |
+
+The issue's expected 492×1280 is contradicted by the actual renderer. Its
+unchanged world viewport rounds to width 488 at max 1280. Applying the real
+renderer's `round(worldWidth * maxSide / worldHeight)` formula gives 1290
+as the smallest integer longest side reaching width 492; only scale changes.
+
+Measured wrap, **not a native renderer verdict**:
+
+```text
+measured wrap: idle 41 frame(s) at 12 fps, f0000 vs f0040: max |d| 212; wrap MAE 0.44359887397323583 vs max adjacent 0.7597497951723703 — PASS smooth wrap; different endpoints
+```
+
+The authored animation lasts 40/12 s; displaying every one of its 41 sampled
+frames gives a 41/12 s file. No last frame was dropped, replaced or blended.
+Infinite repeat metadata is separate from seam quality and endpoint identity.
+
+| Retained output / deleted frame series | SHA-256 |
+| --- | --- |
+| `idle-2x.apng` | `bd31ad5b7d952733417a8ac830c40e73719f95ef0cdb8569e092c2bde3161f0e` |
+| Final 41 PNGs concatenated in filename order | `85e0d85756b3efc384a01a235e38a2e9dc3f178dae626937cd3f511c407ec8d9` |
+| Rejected 488-wide 41 PNGs concatenated | `2fa8b2b905dd788a3653aec38edff10bda4bddc3740439da7803d5e849459dd3` |
+
+The sheet shows all 41 labelled samples with the figure within the viewport;
+that small-sheet verdict does not establish endpoint identity. Both producer
+frame sets (82 PNGs total) were hash-verified and deleted after this verdict.
+
+### Native renderer boundary
+
+**rigc emits no closure verdict; this is the measured wrap**
+
+The original max1290 producer log has no `loop:` line. The 2.68.0
+`rigc render --help` has no loop/closure check flag. Its implementation,
+`src/cli/core_commands.ts:1030–1092`, prints samples, physics and the sidecar
+path; `physicsLine` at 1117–1130 counts stepped constraints, not closure.
+The actual printed line is:
+
+```text
+  ..    physics: 23 constraint(s) stepped at 12 fps
+```
+
+Native closure is **UNAVAILABLE**, not a native PASS or FAIL. A separate
+`rig-parts check` loop definition excludes physics; it is not this renderer's
+output. The first/last max difference of 212 is retained as an observation,
+not promoted into an exact-endpoint acceptance bar.
+
+### Measured optimization comparison
+
+All candidates were encoded on Nova from the original compressed APNG, after
+actual APNG/filter/codec reader preflight and zero/255 error controls. No
+rerender was performed. Independent FFmpeg decode compared every candidate's
+41 RGBA frames with the original 41 decoded renderer frames. Maximum error
+below is the largest absolute difference over all RGBA bytes, on 0–255.
+
+| Encoding | Bytes | Decoded frames / max RGBA error | Actual delays | Decision |
+| --- | --- | --- | --- | --- |
+| Original full-frame APNG | 17,453,151 | 41 / 0 | 41 × 1/12 s | Reference retained remotely |
+| Predictor lossless APNG | **11,351,964** | **41 / 0; every byte identical** | **41 × 1/12 s** | **Selected** |
+| Palette APNG, no dithering | 2,638,904 | 41 / 66 | 41 × 1/12 s | Rejected: not lossless |
+| Palette GIF, no dithering | 2,253,537 | 41 / 66 | 8/100 and 9/100 s | Rejected: error and exact-timing failure |
+
+All three measured candidates are 492×1290 with infinite-repeat metadata.
+Predictor APNG saves 6,101,187 bytes (about 35%) while preserving every RGBA
+byte. The smaller palette candidates are not labelled lossless or silently
+accepted as near-lossless; no nonzero error budget was assumed. GIF's
+centisecond clock cannot express an exact 1/12-second frame delay. The
+receipt retains each candidate's SHA-256, all actual delay fractions and the
+measured FFmpeg commands.
+
+## Verdict
+
+**PASS under HQ's owner-adopted measured-wrap bar** — wrap MAE
+0.44359887397323583 ≤ maximum adjacent step 0.7597497951723703, with width,
+all 41 samples, exact 12-fps delays, infinite-repeat metadata and lossless
+decoded pixels preserved in the selected predictor APNG. Exact endpoint
+equality is not a bar. **rigc emits no closure verdict; this is the measured wrap**.
+
+## Rejected alternatives
+
+- Max 1280: actual width 488 misses the size bar.
+- Upscaling the old recording: would not rerender the accepted build.
+- Dropping frame 40: violates the requested 41-frame preservation.
+- Replacing/blending endpoints, changing motion or suppressing physics:
+  would alter what this accepted build actually renders.
+- GIF: centisecond timing cannot express an exact 1/12-second delay.
+
+## Limits
+
+One accepted office idle only. This proves no new rig, artwork, hand motion,
+retarget, physics settlement or arbitrary-force quality. The rendering includes
+physics; the stored parts-check loop with physics excluded is not a substitute
+for these measured pixels. The APNG keeps the renderer's opaque
+background and every source frame's decoded RGBA values, using lossless predictor
+compression. It remains larger than the rejected quantized recordings. No browser playback or new full toolchain suite was run.