Candidate construction guide · native Resolve not tested

Public archive snapshot update

Both reconstruction paths now hash and parse the same bounded byte snapshot. A concurrent path replacement cannot substitute an unverified core after its checksum check. The core, recipe, SVGs, raw controls and R02 motion repair remain unchanged. This is a toolkit handling fix, not another visual repair.

tooling-snapshot-repair.json records the two helper changes. The original preserved-contract.json stays byte-identical; the current repair verifier checks its original helper pins against the explicit before/after pairs, and still checks every other preserved file against its original pin. README-builder.md preserves the earlier builder description.

R02 correction: stopping capture is a separate event

This derivative repairs one missing relationship in the earlier illustrated workflow. It does not introduce another feature or claim a new source example.

• Local 0–15: capture is active; the finish instruction remains in the composer.
• Local 16–25: finish is requested, capture remains visibly active, and a stopping heading replaces the premature reading/complete labels.
• Local 26–49: the red active marker leaves, a neutral stopped marker appears, and the original completion chip and processing copy return.
• Local 50–167: active leaf content, bounds and alpha are preserved against the original recipe. The log, completion, cards and file rows retain onsets 50,96,111,112. This is numeric/content preservation, not rendered-pixel proof.

The adjacent source frames establish a recorder control present at25 and absent at26. A red-to-neutral marker, a stopped label and the deferred processing text are authored explanations of that event, not measured source artwork or source UI text. No source click, timer, waveform, save toast or easing is invented. Frame26 is the visible control-exit boundary, not a measured backend stop time. The submission remains at16. The original progress track still begins at16; it is explanatory, not recording data or measured task latency.

Both embedded content variants remain meaningful: release procedures versus a fieldwork checklist, with the two original and different SVG badge geometries. The accepted core, source packet, selection and retained PTS are unchanged. The two archive reconstruction helpers have the separately documented bounded snapshot repair; their original and selected hashes are both retained. Canvas1280x720, 168frames, 24000/1001, silent, end exclusive168.

Run the inherited preparation commands below, then:

python3 test_repair.py --project clean/project
python3 resolve_controls.py --recipe clean/project/recipes/recording-to-artifact.json > standalone-fusion.csv
cmp fusion-controls.csv standalone-fusion.csv

No R02 render, native Resolve execution, moving inspection, independent critic acceptance or deployment is claimed. The parent must render serially, then send the actual source first and anonymous outputs second to a different fresh critic. The original camera/scroll/pointer omissions remain. The largest open question is whether this brief stop handoff reads clearly during playback.

The following inherited workflow documentation is updated only where R02 changes state gates or packaging; earlier source observations remain dated records, with the new evidence in correction-source-observations.json.

From a captured session to reusable work

recording-to-artifact / wave05-interface. One independently judgeable mechanism: A completion instruction is submitted, processing states change, and reusable skill cards plus edited files become inspectable evidence of completion.

This is an original code/SVG specimen, not a live product demo. No render, native Resolve validation, independent quality approval or publication is claimed. Quality credit remains zero. The two content variants are inputs to the same mechanism, not two examples.

Start and duration

Read START-HERE.md. Requirements: Python 3.11+ (the pinned handoff uses hashlib.file_digest) and Node.js 22.18+ for the shared core's TypeScript preflight. The python3 commands below must resolve to Python 3.11 or newer. Canvas: 1280 x 720, square pixels, silent. Rate: 24000/1001 fps, never rounded. Duration: 168 frames, exactly 7007/1000 seconds (7.007000000 s). Local frames run from 0 through 167; end is exclusive at 168.

Source: Record & Replay in Codex (2026-06-18), https://www.youtube.com/watch?v=ZK3JhU73W18. Selected source frames [1055, 1223). The source-master hash is inherited, not reverified. Only the retained 960x540 silent excerpt was hashed and decoded here. Its own PTS is retained in retained-excerpt-pts.json; nominal source seconds are frame/rational-rate arithmetic. Absolute source PTS is unavailable. The original source packet and selection unit remain byte-exact in source-packet.json and source-unit.json. The inherited September 5, 2026 cutoff is not refreshed by this work.

Source observations

All 168 consecutive retained-excerpt frames were inspected in native-pixel sheets, with native-pixel crops around boundaries. Large sheets are scaled by the viewing tool; this is static frame inspection, not an uninterrupted moving review or independent critic pass. That inherited observation predates R02. R02 reads the Batch09 R01 review after its own source-first record. An allowed packaging-helper view incidentally returned unrelated legacy review summaries after construction; those were not used. This is not a blind study.

• Local 15, 16: Completion text is in the input through local 15; a submitted bubble and processing response replace it at local 16.
• Local 49, 50, 95, 96: The retained edit jumps to a dense tool log at local 50, then to the completion response at 96. Work-duration text changes across the edit; elapsed task time is not the excerpt duration.
• Local 110, 111, 112: At 111 the view changes to two document cards and an edited-files container; file rows are visible at 112. This is not a card morph out of the recorder.
• Local 144, 167: The final result persists while the source view and pointer continue moving; the source does not end on a perfectly static full-frame freeze.

No easing curve, opacity curve, source typography metric or pixel travel distance was recovered. Those numbers in this recipe are authored estimates.

Authored adaptation

• Original workshop, file names, document badges and two larger review cards replace the source product and demo data.
• Submission 16, log cut 50, completion 96, card cut 111 and file rows 112 are retained as authored event choices on the same clock.
• The illustrated log is shorter; progress tracks are explanatory and do not represent measured latency.
• The last result is held without the source pointer travel or scrolling. Both variants change actual procedure content and vector badge geometry.

Biggest known gap: The source camera/scroll and final pointer motion are omitted; the artifact hierarchy is preserved in a calmer, larger layout.

Frame plan

The following anchors describe this authored build. Compare the separate source observations before calling any anchor a source event measurement.

Mechanism invariants

• One completion-to-deliverables chain; no separate unrelated motion.
• Two reusable cards and a separate edited-files list survive to the last frame.
• Task-time compression remains visibly disclosed throughout.

Remotion contract

Use the unchanged shared Recipe / Content API. No new engine API is needed. The executable recipe is inside active-assets.zip at project/recipes/recording-to-artifact.json. Reconstruction preserves that path. defaults and adaptation are embedded and duplicated exactly in inputs/defaults.json and inputs/adaptation.json. Edit a Content copy, not src/ or tracks.

Text is ASCII, explicitly line-broken, and fixed size. Width preflight checks the complete strings, including a deliberately unrevealed suffix. It does not shrink, auto-wrap or crop rejected copy. Image slots accept the declared SVG aspect and resolution only. The supplied SVGs are geometric primitives with no embedded text, fonts, image data or external references. Their different variant geometries demonstrate asset substitution.

Run these preparation checks from the unpacked kit directory:

python3 test_kit.py --reconstruct clean
python3 test_kit.py --project clean/project
python3 handoff.py verify --guide guide.json --project clean/project
python3 handoff.py frames --guide guide.json --project clean/project > roundtrip.csv
cmp controls.csv roundtrip.csv
python3 test_kit.py --project clean/project --export-fusion > fusion-roundtrip.csv
cmp fusion-controls.csv fusion-roundtrip.csv

CSV files preserve literal helper stdout, including CRLF line endings. Capture stdout as bytes or redirect it directly as shown; never pass it through text-mode newline normalization. Equal parsed fields are not a substitute for a byte-exact cmp. The kit's verifier keeps both raw-byte gates.

test_kit.py --core PATH overrides the default this kit's own core-publication-v08.zip; it never discovers a developer's core implicitly. The same self-contained verifier supplies --reconstruct DEST and --export-fusion, so it does not depend on optional helper filenames. In a parent repack, active-assets.zip is accepted only with the exact overlay checksum and member set. Content JSON may be flattened to defaults.json and adaptation.json; the verifier still requires exact equality to the embedded Content. The nominal bundled-TTF font check is not the browser's final shaping check. The parent must render both variants, inspect readability and motion, and retain the renderer's manifest. Parent-only render commands, after a shared-dependency runtime is prepared, use the existing scripts:

cd clean/project
npm run render -- --recipe recording-to-artifact --variant default --out out/recording-to-artifact-default.mp4
npm run render -- --recipe recording-to-artifact --variant adaptation --out out/recording-to-artifact-adaptation.mp4

No install or encode was run by this worker. Parent batch preparation uses the existing shared dependency loader; do not duplicate node_modules for each kit.

DaVinci Resolve

Follow RESOLVE-HANDOFF.md and resolve-recipe.json. The CSV is a baked per-frame control table, not an automatic importer. It includes evaluated text for both variants, leaf opacity, world bounds and exact normalized units. Group rows are diagnostic, not additional painted nodes. Native Resolve is not available or validated here. Rasterization, font metrics and native rendering agreement remain open even when all numeric checks pass.

Download and review boundary

active-assets.zip is exactly one recipe plus its referenced SVGs under the project/ prefix; overlay-manifest.json pins every member. No dependency tree, source film or source image is in the overlay. download-kit.zip, when supplied by the parent handoff, also contains the exact shared core ZIP, documentation, source metadata/PTS, input JSON, controls and verifier. Source research PNGs stay in the bounded private worker and are never editable assets.

The parent owns all rendering, independent critics and publication. A packaging or test pass is not source-quality approval. This kit does not synchronize an all-hands recording or replace its required storyboard review.