# Worktree census AND REAP — doyle's + todlando's trees on hfenduleam (2026-09-24) ## RESULT (reap executed 17:54:18–17:55:20Z) **51 removed, 17 refused, 95 → 44 worktrees.** Every one of the 17 refusals was `contains modified or untracked files, use --force to delete it`. No `--force` was passed anywhere, so those trees stand: a refusal means unverified content, and the safe direction is to leave it. They are listed in `reap-refused.txt` and want a per-tree read, not a flag. Sizes over the reaped set (not the whole `.worktrees`): apparent **1,683,327,886**, allocated **1,832,480,768**, escaped **0 inodes / 0 B**. This escaped zero is **VACUOUS, not forced**: the population behind it was captured and it is EMPTY — 0 files at nlink>1, 0 distinct multi-link inodes in the entire set. Source checkouts contain no hard links. The clause stays unexercised. **Free-space delta: −5,368,799,232 B — free space FELL by 5.4 GB while the reap removed 1.8 GB.** It is uninterpretable, and the writer census says why: the pre-reap census caught `cargo.exe`, four `rustc.exe` and `Runner.Worker.exe` **in Services session 0** — a CI job (run 36035629802, head `fe81540c`, PR #272's own run) building on this box through the whole window. This is a CENSUSED ambiguity, which is a result; it is not an unexplained gap, and it gets no mechanism invented for it after the fact. **MY PROCEDURAL ERROR, recorded because it is the reusable part: I put the writer census INSIDE the script it was supposed to gate.** The census ran first and printed the runner — but its output only reached me when the whole job finished, minutes after the removals. A census that reports after the action cannot gate the action; it can only annotate it. I had also told two agents to census writers before capturing, and then took my own capture without reading mine. Next time the census is a SEPARATE call whose result is read before the reap is launched. The earlier estimate in this file — "expected reclaim ≈ 7.8 GB" — was **wrong as applied**: it described reaping everything, while the evidence-based set is 68 of 95 trees and measures 1.8 GB allocated. The 27 kept trees hold the balance. Corrected here rather than left standing. Census only. **Nothing was deleted.** The reap is HELD until the volume has a readable writer set (todlando's W6 cargo legs were live during the measurement; hertz's #271 PR run takes the Windows leg after). Captured by doyle, session post-W5-landing, main `2e52677a`. ## Sizes — IR-150's three numbers, whole `.worktrees` tree | number | bytes | command | |---|---|---| | apparent | 7,555,161,221 | `du -sb --apparent-size .worktrees` | | allocated | 8,001,377,280 | `du -sB1 .worktrees` | | escaped (inodes, bytes) | 0, 0 | `find .worktrees -type f -links +1 -printf '%i %n %s\n' \| sort \| uniq -c \| awk '$1 < $3 {n++; b+=$4} END {print n+0, b+0}'` | Escaped 0 here is **not** the forced zero todlando measured on a cargo pool: these trees are source checkouts, not build pools, so the multi-link population is small by construction. The population behind it was not captured; treat the clause as UNEXERCISED, exactly as the three prior zero-escape samples. Expected reclaim of a full reap ≈ allocated − escaped − the live lanes ≈ **7.8 GB**. A free-space delta was NOT taken and must not be reconstructed later: during the window todlando's W6 build was writing the volume (measured 17:0x: 1 `cargo.exe`, 2 `rustc.exe`, 0 `Runner.Worker.exe`), so a whole-volume instrument reads uninterpretable here. Take the delta at reap time, with a fresh writer census, or take none and claim none. ## Build pools: there are almost none 96 worktrees; only **3** carry a `target/` dir, and two of those are empty stubs: - `304-release-note-s4/target` — 0 B - `307-relay-journal-before-emit/target` — 146 B - `w6-divulge/target` — 166,165,244 B and growing — **todlando's LIVE W6 lane** So the reap's value is the checkouts themselves, not artifacts. No `pool-release` is owed by any tree but `w6-divulge`, which is todlando's to release. ## Classification — 96 trees (95 + the main checkout, which is excluded) | class | count | reapable | basis | |---|---|---|---| | DETACHED | 15 | yes | gate rigs and RCA rigs; carry no lane | | MERGED | 30 | yes | branch is an ancestor of main (`git branch --merged main`) | | EQUIV | 23 | yes | `git cherry main ` shows 0 `+` lines — landed by squash/rebase | | REAL | 27 | **no — decide per tree** | `git cherry` shows unlanded commits | 68 of 95 are reapable on evidence today. `git worktree list --porcelain | grep -c prunable` = 0, so every record points at a directory that exists — none of this is pruning, all of it is removal. Ancestry alone would have called 50 trees unlanded; the patch-id arm cut that to 27. Do not skip it — `--merged` cannot see a squash-landed lane, and a tree deleted on that basis is a lane deleted. ### The 27 REAL trees, with their unlanded/equivalent commit counts Two are LIVE and must not be touched: - `feat/331-w6-divulge` (+1) — todlando's W6 lane, activation commit unpushed - `ci/185a-census-test-attribution` (+1) — hertz's #271, open and awaiting its run One is an operator referral, and its tree is evidence: - `docs/golden-head-intake` (+2) — PR #135. The +2 corroborates the standing finding: its content is on main as a SUPERSET, so the merge button would delete 233 ruled lines. Keep the tree until the operator closes the PR. The remaining 24, all uncensused as to intent: ``` build/304-w2-bootstrap-tcp +2 diag/302-meet-offload-field +2 diag/302-net-worker-stamps +2 diag/302-portmapper-off-field +3 ci/traceable-reqs-v0.4 +1 docs/register-locksmith-followups +3 docs/seam-consumers-refusal +1 feat/autostart-emit-carrier +1 -2 feat/emit-single-write +3 -3 fix/206-is-locked-posture +1 -4 fix/208-session-scoped-briefing +1 -2 fix/209-er-inbound-local-bypass +1 -1 fix/74-interior-space-runs +1 fix/ir21-samepkg-bin-edges +1 fix/w3-line-safety-attr +1 -17 golden/307-v071 +1 spec/access-model +13 spec/brain-handoff-pilot +4 spec/brain-handoff-pilot-rebase +7 spec/sealed-code-grill +2 test/fixture-prebuild-hardening +2 test/rig-stop-identity-scrub +1 test/twohost-seal-barrier +2 -7 test/w2-rig-fixups +2 ``` A mixed count (`+1 -17`) means most of the lane landed and a remainder did not — those are the ones worth reading before removal, since the remainder is either a deliberate drop or a lost fix. **Removing a worktree does not delete its branch**, so a tree whose branch is kept loses nothing but the checkout. That is the cheap path for all 24: remove the tree, keep the branch, decide the branch's fate separately and never under disk pressure. ## Procedure when the box is quiet 1. Confirm the writer set: ask every agent whether its leg WRITES this volume (not whether it reads the pool), and check `cargo.exe` / `rustc.exe` / `Runner.Worker.exe` on the box. 2. ~~Expect **IR-153**~~ **DONE 2026-09-24, before the quiet window, because it writes only admin bytes and the quiet window is the scarce resource.** 17 records carried the backslash spelling — exactly the 17 IR-153 left standing as "other agents'", and they are mine and todlando's. `git worktree repair` on all 17 at once: **rc 0, 17 `repair: gitdir incorrect` lines naming exactly its targets, 17/17 verified individually at 0 backslashes, repo-wide rows 17 → 0, and `git worktree list` now shows no `\.git` row.** Predicate controlled both ways before the call: the backslash grep found 17 (positive), and `w6-divulge`'s record read clean forward-slash (negative). Neither live lane was among them. Refinement for IR-153: its batched call exited **1** with 38 stderr lines about records it declined; naming every malformed record and nothing else exits **0** with one line per repair. The per-tree gate stands regardless — the rc is not what was trusted here either. 3. Capture the three sizes per tree before `rm`, then measure the reclaim against allocated − escaped. A gap outside the band is recorded UNEXPLAINED, never given a mechanism after the tree is gone. 4. Reap the worktree subtree only. Move cwd out first (todlando's W5 teardown needed that). `git worktree prune` after. ## Not filed on the board This is infrastructure debt, so it belongs in `docs/INFRA-REGISTER.md`, never the spt-bs-releases board. The register entry is deliberately NOT opened as its own docs PR while two lanes are in flight — it rides the next docs lane, pointing here.