stats

Produce the CoalWash stats report for this project, in the user's language. Tables only, minimal prose. Read-only measurement — the ONE exception is named below, and it is always observed BEFORE it can happen.

First, look at the transaction dir [project]/.claude/coalwash/ directly (a plain listing) for a journal.json — a dangling prior run, and it must be OBSERVED here before the next step can resolve it. Then run the gauge — the one-shot CLI from the engine scripts/lib/: node "[LIB]/cli.mjs" gauge --json (discover + measure + verdict + break-even, PLUS a recovery preflight that writes ONLY if the journal you just checked was there — never hand-compose the lib calls) — plus the state file ~/.claude/projects/<slug>/coalwash/state.json, and show:

• Gauge: current band (LEAN/OBESE/FULL) · always-loaded footprint ~tok/session (label ~est) · Memory-BMI — informational only, footprint/measured-muscle, n/a if muscle is 0 — plus the certain-fat ~tok figure (mechFatFromText: exact-duplicate substance lines + excess blank runs; a lower bound — unread/unprovable content counts as muscle).
• Inherited governance (from the gauge's inherited field, when non-zero): the per-session cost of governance files loaded from ABOVE this project — an umbrella CLAUDE.md and its imports. Show it as its own ~est figure and label it plainly: real context cost this project pays every session, deliberately outside its own footprint, BMI and band, because it is not this project's to wash or externalize. Zero/absent → omit the line. Never present it as reclaimable here.
• Fat reading: the certain-fat ~est tokens measured this gauge (muscle = footprint minus that figure) · when the band is OBESE or FULL, the break-even payback numbers for BOTH halves FULL requires — (a) cutting the certain fat: fat/day at the stamped session rate · one-run cost · break-even days, and (b) reorganizing the RE-TIER envelope's demotable index mass, when one is configured: its own fat/day · break-even days. A store with certain fat but nothing demotable shows (a) only.
• Adjudication rate (board #120 — the one direction nothing measured): false DELETES have two guards (the fidelity gate, the claim-strength outsider); false KEEPS had none, and no surface reported how often this tool actually cuts. Read it from data that already persists, no new state required — [project]/.claude/coalwash/keeps.json (the durable reject log: one {target, reason, date} per adjudicated keep) against the two bin manifests [project]/.claude/coalwash/fat-bin/index.json and .../store.old/index.json (one entry per banked cut, each carrying bytes and an origin of program-cut or wizard-cut). Show: keeps count + the oldest keep's date (a keep never expires and is never re-flagged absent new evidence, so an old one is a decision nobody has revisited) · cut count and total bytes per bin, split by origin · and the resulting rate as N kept / M cut. Counts and byte figures are deterministic — do NOT label them ~est. Name the floor honestly rather than reporting the count as lifetime truth: a bin is retention-bounded, so items swept past their horizon have left the index and the bin count is a LOWER BOUND on cuts ever made; the destroyed ones are in that bin's death.log, so cite it when the rate is being read as history rather than as current state. No keeps.json and no bin dirs → omit the line entirely; a store that has never been adjudicated has no rate, and printing 0 / 0 would read as a finding.
• Lean floor (legacy, historical only): the leanFloorTokens/leanFloorAt fields, if ever stamped — no gauge has read them for the band since task #4 (1.0.0); show value + stamp date, or "never stamped", never as a live threshold.
• Pending ask: an unconsumed band crossing awaiting the next Stop (which band, since when), or none — CoalWash asks once per crossing, never a repeating nag; there is no time-based snooze.
• Sub-spawn bill (the 0o true-bill): from the project's state entry, subSpawns + subParcelTokensAccum — "subs this session: N spawns ≈ X tok parcel (~est)" (every sub spawned from this room re-carried the room's always-loaded parcel; the meter counts silently at the spawn site — main — since hooks never fire inside subs). Zero/absent → OMIT the line entirely, never print "0 spawns".
• Last run: newest snap-[timestamp]/ in the transaction dir (= the last apply's snapshot, with date) and whether the pre-gauge check above found a dangling journal.json (an interrupted run). If it did, read the gauge's recover field and say which it is: resolved (rolled-back / cleaned / no-mutation — the gauge's one write exception did its job, nothing further needed) or UNRESOLVED and kept for you (partial, or none carrying an error — quote the error verbatim; the journal and snapshot are deliberately preserved, and a new /coalwash run writes its own journal at that same path and would overwrite it, so recommend reading it rather than re-running). No transaction dir → no run recorded.
• Class-A estate (P1, report-only — COALWASH_BLUEPRINT.md §19): run node "[LIB]/cli.mjs" estate --json and show total estate size (this project's CC session transcripts + tool-results/subagents/other overflow) · the per-type byte rollup · the heuristic ~est reclaimable figure (entries older than the retention horizon) · machine-wide orphan-slug-dir count + bytes if any (a slug whose owning project path no longer exists — candidates, not confirmed). Byte figures are deterministic; the reclaimable figure is ~est, always labeled. This is MEASUREMENT ONLY — nothing is deleted, archived, or edited; name the honest ceiling verbatim: "P1 = report-only; P2 (retention/archive/restore) runs only through the estate wizard commands, never this read-only path." No estate found → omit the section entirely, never print a zero-line.