feat(floor): cost.json shows stage elapsed time and deterministic work per run (6.35.0)

[PrzemekGalarowicz]

Summary

A maintainer can now answer, for one concrete /pharn-loop or /pharn-ship run, where the model usage went, where the observed wall-clock time went, which expensive deterministic operations ran, and which were avoided by the 6.33.0 / 6.34.0 reuse mechanisms — from cost.json, with no second telemetry system.

• executions (a VIEW, no new instrumentation): derived from the existing phase markers (pharn/floor/stage-executions-core.mjs, method stage-start-to-return/1). Every current-run stage-start is one row, ended only by the next marker when that marker is the orchestrator's return. Re-runs are run 2, never merged. Anything else is unmeasured with a closed reason and elapsed_ms: null — never 0, never a guessed end. It is labelled everywhere as observed wall clock, not CPU, model or tool time, and not monotonic.
• work[] (FACTS captured at the moment): /pharn-regress and /pharn-verify each append one line at done to .pharn/cost/<feature>/work.jsonl (pharn/floor/stage-work.mjs). The counts come from the gate-run stamp the verdict used: executed / reused / nothing-to-run / required, BASE evidence fresh|reused, and the install's exit and ms, which is the only new timer. Per-iteration capture is needed because both reports are overwritten every iteration. The append is observational and best-effort: no exit, verdict, reuse, route or commit depends on it.
• Checker rule 9: validates work[] rows and their run-window membership, and recomputes executions from the file's own facts. /2 admits exactly the current key set or the pre-6.35.0 one. No schema bump.
• Run report and stop table: three separate blocks — tokens, elapsed, work. Nothing is subtracted from anything else.
• Unchanged: mark-phase.mjs and its printed binding line, run membership, context binding, dedup, stage attribution (✧ equality test), and --verify-transcript.

SKILLS_VERSION 6.34.0 → 6.35.0. MIN_CLI stays 0.5.0. Contract: pharn/pharn-contracts/cost-ledger.md, "Stage executions and deterministic work".

Evidence

• .dev/features/run-performance-breakdown/DEMO.md — a controlled end-to-end fixture (not a live run). It uses the real marker writer, work-record writer, emitter, checker and run report, and answers the five questions from cost.json. It also reports the measured overhead: under 1 ms added to render and check on a 126-request run, one extra small file read, no transcript pass, no process, no model call.
• The real stage scripts are exercised in stage-regress.test.mjs (★ HIT, the budgeted chain) and stage-verify.test.mjs (★ EQUIVALENCE, OBSERVATIONAL). Those tests assert the work counts against the processes the fixtures counted spawning.

Pipeline verdicts (/pharn-dev-ship)

• plan lessons check: exit 0. validate: GREEN.
• regress: no-regressions. The first regress attempt was void — it read another session's stale result files in shared .pharn/ scratch. It was redone from run-private scratch; see REGRESSION.md, "A void first run" (promoted as L66).
• verify: raw FAIL on lint:md only. The single offending file is .pharn/pr-body.md, another session's gitignored scratch file, not part of this PR. Measured on a clean copy of the tree, lint:md exits 0 and the verdict over the same map is PASS; VERIFY.md carries both. npm test: 4502/4502. Reconcile: CLEAN.
• Independent opus review (REVIEW.md): nine findings. Two were major: a FIFO planted at work.jsonl could hang a stage before done, and an unknown run window read as "no work". All are fixed with tests, or stated as bounds, before verify ran.
• GATE 1 was delegated to the model by the maintainer's prompt; the merge decision is the maintainer's.

Named residuals

gate-process-duration (no per-gate timer), ship-spec-elapsed (/pharn-ship marks no spec stage-start), work-on-non-done-exit, work-record-provenance.

🤖 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: 50979471-2494-4f41-bb98-35f78103e78a

---

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.}