reveng-js-recovery-climb

name

reveng-js-recovery-climb

description

Use when climbing JS recovery / unlockable coverage on Claude SEA or similar minified bundles, editing js_recovery_toolkit (Waves 7–10: ensemble, structural / Bun map, defrag / word_map, normalize / semantic / LLM, webcrack + tag-boost, Hungarian soft_assign + unique-token tombstones), running scripts/js_recovery_toolkit.py, writing test_wave8_* / test_wave85_* / test_wave9* / test_wave10_*, or reporting oracle_coverage vs survivor/unlockable ship-bar metrics. Enforces Option C honesty (100% unlockable ship bar; oracle aspirational; report BOTH) and refuses GA / R-RALPH-2 / exe-decode overclaims.

REVENG JS recovery climb (Option C)

When this applies

Any work on src/reveng/app_reverse_engineering/js_recovery_toolkit/, the scripts/js_recovery_toolkit.py CLI, wave-tagged unit tests, Claude SEA / Bun .map recovery dogfood, or status language that mentions oracle vs unlockable coverage.

Option C ship bar (non-negotiable)

Metric	Role
Survivor / unlockable coverage	Ship bar = 1.0 (100%)
oracle_coverage	Aspirational — climb is real; never the sole merge claim
recoverable_oracle_coverage	Climb vs map minus unique-token tombstones (Wave 10)
Reporting	Always report BOTH unlockable and oracle (plus recoverable when tombstones run)
Wave 10 status	Soft-assign + tombstones merged PR #150 → 82bc0ec3 (Sol PASS tip2 accf553a). Closeout CEO: ceo-update-2026-08-21-wave10-closeout.md.

Do not invent metrics beyond pinned dogfood (see references/pins.md).

100% ceiling honesty (CEO Q&A — do not regress)

Claim	Truth
Unlockable / survivor 100%	Already the ship bar — achieved on Claude Bun dogfood by construction once modules unlock
Full-oracle 100% on this stale-map → Claude SEA (sourcemap_size=0)	Not honestly possible for ~874 unique-token tombstones (no residue in the Bun blob)
Full-oracle 100% in general	Yes when a same-era .map exists (already measured 1.0) or Bun embeds SerializedSourceMap (sourcemap_size > 0)
Unique-residue survivors	~82% recoverable; ~18% still climbable with better algorithms — not guaranteed to 100% without false positives
“Pretty close” / ~57% oracle	≠ 100% — never redefine near-miss as done

Refuse status language that collapses these into one “100% recovery” claim.

Wave map (do not renumber casually)

Wave	Focus
7	Ensemble signals / index
8	Structural match, Bun serialized sourcemap, singleton coverage
8.5	Iterative defrag + word_map
9	Readable normalize + semantic digest + LLM digest
9b	webcrack + LLM tag-boost (provider_summarize / tag_boost)
10	Hungarian soft_assign (scipy.optimize.linear_sum_assignment) + unique-token tombstones + margin gate

Merged lineage: PR #149 = Waves 8–9b; PR #150 = Wave 10 (open/merge as of skill authoring).

Algorithms in play

• TF-IDF cosine similarity (word_map / soft_assign word channel)
• Char n-gram TF-IDF (fixtures / small bundles; skipped on multi-MB Claude cli.js)
• Greedy unique assignment → replaced/supplemented by Hungarian for collision traps
• Unique-token tombstones (shared boilerplate ≠ survivor)
• Margin gate (reject ambiguous best≈second pairs)
• Graph unlock / co-occurrence / semantic_digest (defrag loop)

Inspiration (cite, do not overclaim identity): Sinkhorn/assignment arXiv:1707.07285; unsupervised clone measures arXiv:2401.09885. Pins + access dates in references/pins.md.

Non-negotiables checklist

JS recovery climb:
- [ ] Ship claim uses unlockable/survivor coverage (bar = 1.0), not oracle alone
- [ ] oracle_coverage reported alongside unlockable (BOTH) — aspirational only
- [ ] Numbers match pins.md or a NEW dated dogfood receipt (no invented deltas)
- [ ] Hermetic tests: test_wave8_* / test_wave85_* / test_wave9* / test_wave10_* green for touched wave
- [ ] CLI exercised via scripts/js_recovery_toolkit.py when behavior changes
- [ ] Dogfood receipts stay under /mnt/c/tmp/reveng_w* — NEVER git-add Anthropic trees / Bun extracts
- [ ] Honesty: NOT R-RALPH-2 close, NOT exe decode, NOT enterprise GA
- [ ] Named-path commit only (reveng-named-path-commit); verify git diff --cached --name-status
- [ ] Same-era .map materialize ≠ Claude SEA path (sourcemap_size=0) — do not conflate
- [ ] Tombstone scan uses unique-owner tokens (any-token hits false-survivors on 24MB bundles)

Local surfaces

• Package: src/reveng/app_reverse_engineering/js_recovery_toolkit/
• CLI: python scripts/js_recovery_toolkit.py (prefer /usr/bin/python3.9)
• Tests: tests/unit/test_wave8_*.py, test_wave85_*.py, test_wave9*.py, test_wave10_*.py
• Fixtures: test_samples/js_recovery_toolkit/
• CEO (closeout, plain English): docs/architecture/ceo-update-2026-08-21-wave10-closeout.md
• CEO (feature metrics): docs/architecture/ceo-update-2026-08-10-wave10.md
• Research: docs/architecture/research-wave10-soft-assignment-2026-08-10.md (+ wave8/85/9*)

Climb loop (agent)

1. Research — Read pins; cite arXiv with access date; Exa if up else WebSearch + L44.
2. Hermetic TDD — Fail-first on the wave under edit; keep tests offline / fixture-bound.
3. Dogfood — Operator-local Claude SEA (or agreed target); receipts only under /mnt/c/tmp/reveng_w*.
4. Honesty docs — Update status with BOTH metrics; cross-check reveng-release-honesty.
5. Named-path PR — Stage exact toolkit / test / doc paths; never dogfood trees.

Anti-patterns

Excuse	Reality
"oracle ~57% — almost GA"	Oracle is aspirational; ship bar is unlockable 1.0. Not enterprise GA.
"Map path got oracle 1.0, SEA is done"	Same-era .map materialize ≠ Claude SEA (sourcemap_size=0).
"Recoverable ~82% of residue = unlocked all"	Residue recoverable ≠ unlockable ship bar; report BOTH + tombstone count.
"Commit /mnt/c/tmp/reveng_w* for reproducibility"	Dogfood is operator-local, never in git.
"Wave 10 closes R-RALPH-2 / exe decode"	Toolkit climb ≠ R-RALPH-2 close ≠ native exe decode.
"Skip unlockable; oracle moved +3pp"	Option C: unlockable 100% is the bar; oracle climb is bonus.
"git add -A after dogfood"	Named-path only; tmp receipts must stay unstaged.
"Any token in bundle ⇒ survivor"	Use unique-owner tokens or mega-bundles mark everything alive.
"57% oracle / 82% recoverable ≈ 100%"	Pretty close ≠ 100%. Tombstones block full-oracle on SEA.
"Smarter algo will unlock tombstones"	No unique residue ⇒ not an algorithm miss; need map/embed or accept ceiling.

Measured dogfood (Claude SEA, operator-local — not in git)

Pinned snapshot (see pins for dates):

• Oracle climb: ~29% → 33% → 52% → 55% → 57% (1087/1902)
• Recoverable ≈ 82% of unique-residue survivors
• Tombstones: 874
• Unlockable: 1.0
• Contrast: same-era .map materialize → oracle 1.0 when map exists; Claude SEA sourcemap_size=0

Cross-refs

• Release / GA honesty: reveng-release-honesty
• Commits on dirty DrvFS: reveng-named-path-commit
• Pins + paper URLs: references/pins.md
• Dynamic workflow: ~/.claude/workflows/reveng-js-recovery-climb.md
• Honesty closeout (different loop): ~/.claude/workflows/reveng-wave-honesty-closeout.md