Revisit densify: since-last-look protocol

[24601]

Fresh PR off latest main (c82e066 / merged #44). Does not amend #44. Does not bump the 0.5.0 pin. invented_signal: false. Protocol, not a census fold.

What this is

Basit standing order: Augustus revisits already-catalogued repos when they change since last look, not only first sightings. Things change a lot.

Hourly fold prompts must treat revisit HIGH like novel HIGH. Skip is for star-noise and name collisions, not for "we already have a card."

1. Fingerprints to store so hourly can diff: default_sha, pushed_at, description_hash, release_tag. Store: research/revisit_fingerprints.json (seeded from SemIf §117 and NanoJev §115, not a live census this pass). Helper: research/revisit_fingerprints.py (offline, no network). SHA move is not a replica. The classifier never invents equivalence.
2. Material vs star-noise. Material (revisit HIGH): README rewrite, API / primitive surface, release tag, calibration claim, serving port or bottle, bench rewrite. Star-noise (pulse only): stars / likes / forks alone. A star jump on a thin README is still star-noise. A SHA move on a 0-star repo is still material.
3. Densify cards update prior notes. Keep the original notes.md section id. Append a dated since-last-look card. Do not mint a sibling first sighting. Quote the new README / API / release theirs. Keep the prior quotes. Namesake locks stay. Prior uniqueness locks stay one consecutive substring.
4. Hourly fold prompts. Paste block lives in research/revisit-checklist.md. CONTRIBUTING and SKILL step 8 / REVISIT / since last look point the same way. scripts/hourly-refresh.sh runs the helper --self-test.

IDs: notes.md §122. Merged #44 owns §121 / items 369–384 / batch #104. Unused 317–321 stay unused. No composition items 385+.

Adversarial checklist (reviewer hooks)

• Invented metrics. None this PR. Store is seeded, not fetched. Helper --self-test does not hit the network. invented_signal: false.
• Soft ≠ hard. Star-noise is not a fold. SHA move is not a replica. Wire-compat is not a calibrated Noul. A new bench number is not Harbor.
• Densify vs new. Do not mint a sibling first-sighting section for the same source. Rename densify is still densify (SemIf §117).
• Keep-both walls. 0843 / 0915 / jcr / 0922 / 0940 / 0947 / 1049 / 1143 stay one consecutive substring each across 21 overlays. Do not mutate them. Do not reopen or amend PR Fold hourly 0042 HIGH: jev-readout, jevify, laya-jev-lab #23–Fold hourly 1143 HIGH #44.
• CHANGELOG hygiene. Unreleased stays scannable. uniqueness_gate fails if CHANGELOG holds the revisit lock dump. Protocol lock lives in SKILL.md + research/ + CONTRIBUTING (5 files), not a 21-overlay dump wall.
• Version pin. SKILL metadata, marketplace, CITATION stay 0.5.0.
• Anti-patterns to refuse. Treating a catalogued card as done. Folding stars/likes/forks. Inventing equivalence from a SHA move. Copying keys / npm / pip / uv / docker.
• Section IDs. Do not renumber §114–§121 / items 289–384 / batches #97–#104.

Tests

python3 research/revisit_fingerprints.py --self-test
python3 .agents/skills/augustus/scripts/uniqueness_gate.py
python3 .agents/skills/augustus/scripts/evaluate_decisions.py --self-test

All three pass on this revision. uniqueness_gate checks 0843 + 0915 + jcr + 0922 + 0940 + 0947 + 1049 + 1143 as one substring each across 21 overlays, plus the revisit lock as one substring across 5 protocol files. CHANGELOG does not hold the revisit lock dump. Helper grades unchanged / star_noise / material and always returns invents_equivalence: false.

Parent merge only after CLEAN adversarial review AND testing (Basit standing order). Do not merge this PR from the agent.

  

[24601]

Testing gate: PASS

Independent adversarial review + tests on head 7065a3d vs origin/main c82e066 (merged #44). GitHub: MERGEABLE / CLEAN. Pages build SUCCESS (run 35529591401). Do not merge from this agent.

Pushable FAIL found on 1011dd5 and fixed in 7065a3d: seed store invented description_hash a3c0f1b9e2d4 (not sha256[:12] of any quoted description), put NanoJev README SHA prefix in description_hash, truncated SemIf HEAD to 12 chars, and used commit time 04:46:33Z as pushed_at (notes §117 push clock is 04:46:36Z). That would make the first honest hourly compare always look like a fingerprint move, and it mixed README SHA into the description field. Store now uses full §117 HEAD, the GitHub push clock, description_hash: null when notes do not quote the GitHub/Space description, and README SHA as an optional extra. Helper --self-test rejects truncated SHAs and README-as-description_hash.

#	Check	Result	Evidence
1	Protocol: material vs star-noise; densify prior notes; no sibling first-sighting; no invented equivalence	PASS	SKILL step 8 + ## REVISIT / since last look; checklist densify rules; helper invents_equivalence always false; SHA move is material / revisit_high_like_novel_high, not a replica. Stars/likes/forks alone are star_noise / pulse_only.
2	Fingerprints default_sha, pushed_at, description_hash, release_tag; helper --self-test	PASS	Store fields locked. SemIf SHA ca3ba65f142967030ecb453346e94d6f476a69df, pushed_at 2026-09-19T04:46:36Z, description_hash null. NanoJev SHA full, pushed_at null, release_tag unified-games-v1, description_hash null, readme_sha extra. python3 research/revisit_fingerprints.py --self-test → revisit-fingerprints self-test ok.
3	uniqueness_gate 0843+0915+jcr+0922+0940+0947+1049+1143; revisit lock is protocol substring not CHANGELOG dump	PASS	Eight UNIQ_* strings byte-identical vs c82e066. Gate: 21 overlays; revisit lock 447 chars on 5 files (SKILL, notes, research/README, checklist, CONTRIBUTING). CHANGELOG does not hold the lock dump.
4	evaluate_decisions.py --self-test; SKILL stays 0.5.0	PASS	Evaluator unchanged vs main; self-test ok. SKILL metadata.version: 0.5.0; marketplace / CITATION / README Current / Pages kicker v0.5.0. No release bump.
5	Human-facing prose; hourly fold prompts treat revisit HIGH like novel HIGH	PASS	New prose has no em-dash walls. Checklist Paste into hourly fold prompts points revisit HIGH at novel HIGH. SKILL step 8, CONTRIBUTING, hourly-refresh.sh comment, refresh-jev-research.sh revisit line match.
6	No false Archer landing; no release bump	PASS	New helper/checklist/json have no Archer landing. Refresh-log / changelog-hourly skip Archer. 0.5.0 pin held. invented_signal: false.

IDs. notes.md §122 protocol only. Merged #44 keeps §121 / items 369–384 / batch #104. Unused 317–321 still absent. No composition 385+.

Adversarial (fixed, then PASS)

• Seed honesty as above (7065a3d).
• Densify vs new: protocol forbids sibling first-sighting; rename densify stays densify (SemIf §117).
• Do not reopen Fold hourly 0042 HIGH: jev-readout, jevify, laya-jev-lab #23–Fold hourly 1143 HIGH #44. No wrappers / keys / npm / pip / docker recipes in the protocol prose.

Notes, not FAILs. Cron hourly-refresh.sh still only clones NEW names; fingerprint classify is the fold-agent path (checklist paste), not a live GitHub client. Null description_hash this pass means the first hourly that hashes a live GitHub/Space description will grade material (fail-closed to revisit HIGH), then persist the real digest.

revisit-fingerprints self-test ok
uniqueness-gate ok
0843 chars=2423 0915 chars=1132 jcr chars=1499 lock0922 chars=1292 0940 chars=1350 0947 chars=2350 1049 chars=2162 1143 chars=2188 revisit chars=447 overlays=21 revisit_overlays=5
self-test ok

Ready for parent review. PR stays open. Do not merge from this agent.