feat(floor): /pharn-verify reuses equivalent REGRESS/HEAD gate executions within one delivery run (6.34.0)

[PrzemekGalarowicz]

Summary

Inside one /pharn-loop or /pharn-ship run, /pharn-verify now records a gate's result from the completed
execution /pharn-regress's HEAD side already made of it — without spawning the gate again — when tested code proves
every bound input is equal. Every other gate runs exactly as before; missing or unusable reuse evidence is never a
failure. SKILLS_VERSION 6.33.0 → 6.34.0 (minor). MIN_CLI stays 0.5.0.

• Execution identity (pharn/floor/gate-reuse-core.mjs, pure): sha256 over command (shell/argv), ordered files,
  cwd realpath, --timeout-ms, git HEAD, the PHARN-added env var name, and the tree fingerprint (algo + fp_before).
  Recorded on every run as the additive runs[].identity_sha256.
• Eligible: completed process exit 0..125 (a completed RED is reused like a green), not timed out, did not move the
  tree, no regular results file. Never reused: AC level gates (test, test:e2e, e2e), style gates (they read
  fingerprint-excluded REGRESSION.md/VERIFY.md), reconcile. In practice: typecheck, type-check, build.
• Source binding (pharn/floor/head-reuse-offer.mjs): regress publishes a git-dir offer bound to the HEAD stamp's
  bytes and the open run marker, and discards it at its fresh start; a stamp planted under .pharn/ (e.g. a --quick
  build) is never offered.
• Truthful evidence: a reused entry is ran: false, reason: "reused", with a reused block naming the source
  execution; logs are copied only after hash verification. verify-report.json gains the additive, advisory
  gate_reuse block; VERIFY.md says "reused, not re-executed". Verdict, AC gate, completeness, check-verify,
  check-regress, check-loop-fresh and drain/budget/resume are unchanged.

Measured (.dev/features/verify-head-gate-reuse/MEASUREMENT.md)

Fixture with test, lint, typecheck, build (1 s gates), 3 runs each:

	verify processes	total gate processes	verify wall-clock
before (main)	4	12	~4.8 s
after	2 (+2 reused)	10	~2.5 s (−49%)

Regress unchanged; the offer decision costs ~17 ms. No token saving claimed.

Validation

• /pharn-dev-regress: no-regressions (base 2cf0e85).
• /pharn-dev-verify: PASS — test 4455/4455, validate, lint, format:check, lint:md, reconcile all 0.
• Mandatory fresh-vs-reuse equivalence test, scenarios A–H through the real stage scripts with a spawn counter,
  per-component negative controls (unit and runner level, each mutation-checked), crash/resume, ★ HOOK test on the
  offer path, ★ WIRING with the committed command lines.
• Independent grill and review (read-only Opus agents): no false-HIT path found; all findings fixed
  (GRILL.md, REVIEW.md).

For the maintainer (human-only)

• LIMITS.md "/pharn-verify re-runs the project's own gates" is now partly stale — not edited (human-only).
• A separate git dir under a temp root, in an installed project with no run open, is writable by the Write tool; this
  bounds both the new offer record and the 6.33.0 BASE record. Named, not fixed.
• GATE 1 and the GATE-2 fixes were delegated model decisions (recorded in SHIP.md), not human approvals.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 68069c36-d5b5-41f9-9e04-e0b7008498cc

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}