REPORT — phase6-janitor-window-review
STEP: phase6-janitor-window (Phase 6: the janitor pin stops scanning the shared tempdir)
REV: 498cce3   FINAL: 819ddde3796ac13639a2be017bbe7f7342e520b8
CHANGED: docs/research/phase6-janitor-window-review.md — this review report; no test file was added to the tree: every seam the review brief's REVIEW line names was attacked on scratch copies of REV/FINAL (the pin `cmp`-identical to its `git show` pre-image, the mutants applied to the scratch's `src/orchestrator/oracle.py`, the shared `.venv` symlinked in, `pytest.ini`'s `pythonpath = src` driving the scratch's own sources) and by gate re-drives whose printed lines are recorded below, so the guarantees the review attacked were re-derived live in VERIFIED instead of riding new nodes.
VERIFIED: REVIEW ACCEPTANCE, run first — `scripts/check-scope 498cce3 819ddde docs/research/phase6-janitor-window-brief.md` — exit 0, prints `claimed: tests/test_probe_oracle.py scripts/check-tests`, `tripped: none`, `scope: ok (3 path(s) in 498cce3..819ddde)`. The three paths are exactly the brief (`docs/research/phase6-janitor-window-brief.md`, added at `e9a37db`), the pin (`tests/test_probe_oracle.py`), and the report (`docs/research/phase6-janitor-window-implementer.md`); `scripts/check-tests` is untouched across the range (`git diff 498cce3 819ddde -- scripts/check-tests` empty), so the delivered reading is the brief's preferred re-scope and the serial-set alternative is correctly not taken.
VERIFIED: the embedded copy of the brief under review is the committed brief — the extraction between the BEGIN/END markers diffed against `docs/research/phase6-janitor-window-brief.md` at HEAD `819ddde` is empty (`diff` exit 0; both 603 lines), so the review runs against the committed file and no ORIGIN: BRIEF divergence is file-copy noise.
VERIFIED: the implementer's first acceptance form re-ran green in the documented shape: `.venv/bin/python -m pytest tests/test_probe_oracle.py -q` → `16 passed in 1.97s` (the report's own line `16 passed in 2.01s`; the brief's baseline `16 passed in 1.96s` — same count, the wall is this drive's). `tests/test_probe_oracle.py -q -n 6` → `16 passed in 2.05s` (the report's `2.03s`) — the module is green when spread over workers. `tests/test_probe_oracle.py --durations=5 -q` → `16 passed in 1.95s`, slowest five `0.51s` `test_harness_deadline_beats_the_probe_budget_review_n3`, `0.50s` `test_probe_deadline_and_base_exception_are_contained`, `0.26s` `test_candidate_mutation_of_its_graded_copy_is_a_fail`, `0.22s` `test_flooding_candidate_is_bounded_and_graded`, `0.19s` `test_expected_values_never_reach_the_candidate_process` — the same five members the report's line names (its own figures `0.52`/`0.50`/`0.27`/`0.19`/`0.17`), the janitor pin below the five in both drives; every figure above is the command's own printed line (rule 14).
VERIFIED: `scripts/check-tests --fast` (the implementer's second acceptance command, documented form) — exit 0, `984 passed, 110 deselected in 88.53s (0:01:28)` — the same counts as the report's own line (`984 passed, 110 deselected in 90.26s`) and the brief's EXPECTED WALLS line (`984 passed, 110 deselected in 89.73s`); the wall is this host's, a baseline not a bound.
VERIFIED: the interleave — the form where the retired F2 red lived — `scripts/check-tests --fast --parallel` driven twice by this review at HEAD: parallel phase `950 passed in 21.25s` then `950 passed in 20.34s` (`created: 6/6 workers`, `--dist worksteal`, the janitor pin inside the pool — `tests/test_probe_oracle.py` is not in `SERIAL_MODULES`), serial phase `34 passed in 12.26s` then `12.25s` (batch3 + milestone3), exit 0 both runs; the report's hook figures (`950 passed in 22.17s` / `34 passed in 12.58s`) reproduce as counts, and `950 + 34 = 984` — the member did not redden in any drive.
VERIFIED: `scripts/check-tests` (the full standing battery, the documented form, this review's falsification ground) — exit 0, `1094 passed in 1152.58s (0:19:12)` — the same 1094-test population the exit-gate-runs chain's standing-battery lines display, green on this review's own drive; no instrumentation flag was added (adding one would substitute a non-documented form).
VERIFIED: `scripts/check-static` — exit 0; gated `All checks passed!`; advisory `Found 104 errors.` (unchanged from the brief's baseline at REV and the report's own line; the pin edit draws no new advisory). `scripts/check-docs` — exit 0, `🔍 410 Total (in 809ms) 🔗 183 Unique ✅ 410 OK 🚫 0 Errors` — identical counts to the report's pasted line (`410 Total (in 744ms) …`) and the brief's EXPECTED WALLS (`410 Total / 183 Unique / 410 OK / 0 Errors`); the count's stability across the range is consistent with the two added research files carrying no internal lychee links (the same F28/F33-class bookkeeping the carried notes own — the brief's baseline figure reproduces exactly).
VERIFIED: the implementer report's scope line reproduced — `scripts/check-scope 498cce3 a14afb2bd67002e154661daafb7383794ffa8fc3 docs/research/phase6-janitor-window-brief.md` — exit 0, `claimed: tests/test_probe_oracle.py scripts/check-tests`, `tripped: none`, `scope: ok (2 path(s) in 498cce3..a14afb2)` — matching the report's own VERIFIED line (the range's second path is the brief the range includes at `e9a37db`).
VERIFIED: the pin's change is exactly the report's CHANGED line — `git diff 498cce3 819ddde -- tests/test_probe_oracle.py`: `_probe_tree_names()` → `_probe_tree_names(root)`, the pin takes `monkeypatch` and redirects `tempfile.gettempdir` to `tmp_path`, and both the `before` capture and the assertion scan the run's own root with the assertion at `:178`. Whole-`gettempdir()` coupling survives nowhere in the window scan — `grep -n gettempdir tests/test_probe_oracle.py` names only `:159`/`:161` (docstring) and `:166` (the redirect); `grep -n` over `listdir`/`scandir`/`iterdir`/`glob` names only the `:143` listing inside the root-parametrised helper; and `_probe_tree_names` is invoked only at `:170`/`:178`, both with the run's own tmp root — attacked-and-not-falsified.
VERIFIED: red at REV re-driven on the scratch `/tmp/kilo/jw-rev` (`git archive 498cce3 | tar -x`; the pin `cmp`-identical to `git show 498cce3:tests/test_probe_oracle.py`; the shared `.venv` symlinked in; mutants applied to the scratch `src/orchestrator/oracle.py`). M1 (`725,726d724` — both `_register_probe_tree` calls dropped; the mutant's whole diff vs `git show 498cce3:src/orchestrator/oracle.py` is exactly `725,726d724` on the two registration lines): pin node → `1 failed in 0.16s`, named failing assertion `tests/test_probe_oracle.py:170` — `AssertionError: assert {'orchestrator-probe-exec-mdworkuj', 'orchestrator-probe-exec-mdworkuj-root'} == set()`, the run's own trees standing in the SHARED tempdir, reaped by this review after the drive. M2 (oracle restored byte-pristine, then `726d725`; the mutant's whole diff is exactly the one sandbox-root registration line): node → `1 failed in 0.15s`, same `:170` assertion — `{'orchestrator-probe-exec-jvtek56p-root'}` standing, reaped after the drive. Both the report's REV figures (`1 failed in 0.15s` / `1 failed in 0.17s`) and the standing names' class reproduce; across all four of this review's scratch drives the mutant diffs are exactly the janitor review's M1/M2.
VERIFIED: red at FINAL re-driven on the scratch `/tmp/kilo/jw-final` (`git archive 819ddde | tar -x`; the pin `cmp`-identical to `git show a14afb2:tests/test_probe_oracle.py`; same mutants, same diff-verification against `git show a14afb2:src/orchestrator/oracle.py`). M1 → `1 failed in 0.14s`, named failing assertion `tests/test_probe_oracle.py:178`, the execution copy and its sandbox root standing; M2 → `1 failed in 0.13s`, same successor assertion — `{'orchestrator-probe-exec-ayk48egb-root'}` standing. After both drives `ls -d /tmp/orchestrator-probe-*` printed nothing: the mutant leak stays INSIDE the run's own tmp root (bounded by pytest's own tmp retention), the shared tempdir is untouched — the brief's FOCUS leak clause holds, attacked-and-not-falsified.
VERIFIED: the retired coupling class attacked directly, made deterministic without the battery: a bounded out-of-process creator (a python loop of `tempfile.mkdtemp` over both `orchestrator-probe-` prefixes at full cadence, driven from the review session's shell out of process) hammering the shared tempdir while the pin node drove 6× on each scratch. At REV: three `1 failed` drives reddening `:170` with mid-window peer names in the set-difference (and greens when the cadence straddled the window), reproducing the closeout-status review's F2 red class byte-for-byte in kind. At FINAL: six `1 passed` drives (`0.13–0.19s`), including one while the creator's own churn byte-exhausted the tmpfs mid-window (`Errno 28` raised by the creator itself, not the pin) and the pin stayed green. The whole-`gettempdir()` peer-coupled red class is destroyed by the re-scoped window — attacked-and-not-falsified (the creator's residue was reaped after the drives: the shared /tmp ends at zero `orchestrator-probe-*` entries; the unbounded first creator's pileup of ≈449k timeout dirs was this session's own residue, deleted session-side, no repo effect).
VERIFIED: the standing-before-window facet — two manually planted `/tmp/orchestrator-probe-*` trees standing across the full module at REV and at FINAL: `16 passed` both — a peer tree standing before the `before` capture is inside the before-snapshot of both the old and the new window, so the coupling was always the mid-window creation class the creator drives; both facets are covered.
VERIFIED: the brief's RESIDUALS rule-3 selection reproduces — the accepted triggers in `docs/residuals.md` name only `src/orchestrator/planning_pilot.py`, `controller.py`, `reducer.py`, `decomposition.py`, `model.py`, `derive_obligations.py` and `orchestrator/` paths (grep over the committed file at HEAD), none matching `tests/test_probe_oracle.py` or `scripts/check-tests`; and the report's `TRIPPED: none` cell is the gate's own `tripped: none` line both check-scope drives print.
VERIFIED: the trailer — `git log --format` over `498cce3..819ddde` shows `a14afb2` and `819ddde` ending with `Actor: IMPLEMENTER` (the discipline the exit-gate-runs review's F4 asks for, restored by this step's own commits), and `e9a37db` (the planner's brief) ending without one — see the carried F4 note. This review authored none of the commits under review: its session began at `69ef321` (the review brief) already HEAD, and both commits pre-date it.
VERIFIED: the record's ledger reconciles the step's received set — the closeout row closes its carried set as actionless and the row's own F2 cell (the janitor pin's scan coupling) names a later brief's re-scope or serial-set move as its next action; the later brief (`e9a37db`) took the re-scope reading and this step's implementer deliverable (its own F2, fixed with red evidence) is that absorption, re-driven by this review at REV and FINAL above.
NOT VERIFIED: none
FINDINGS: none — every seam the brief's REVIEW line names was attacked (the mutant reds at both revisions, the whole-`gettempdir()` coupling survivor scan, the isolation and `-n 6` modules, the two interleave drives plus the full standing battery, the scope lines, the copy fidelity, the REPORT forms) and every one held; the figures this step's report carries are reproduced by this review's own drives above. Attacked-and-not-falsified: the hardened pin, its red class, and its independence from peer state; the brief's fix readings, EXPECTED WALLS and RED-evidence account. No failing probe needs registering in `tests/open_defects.yaml` and no `test_review_*` probe is added.
SHAPES: three probed, unbuilt.

- The mid-window peer-created-tree coupling shape (the retired F2 red class): peer trees created by an out-of-process creator inside the pin's set-difference window, probed by 6× pin drives per scratch under the hammer (REV 3 red `:170` / FINAL 6 green, one green under tmpfs byte-exhaustion), VERIFIED above. Not built as a test: the class is retired — its FINAL greenness is the fix itself and the pin's own-root red class is what the pin grades; a committed reproducer would re-choreograph cross-process timing, not pin delivered code (the closeout review's same justification).

- The standing-before-window peer-tree shape (the coupling's other facet): planted `/tmp/orchestrator-probe-*` trees across the full module at REV and FINAL, `16 passed` both, VERIFIED above. Not built as a test: no committed member owns that facet — the before-snapshot absorbs the standing shapes in either window by construction.

- The four scratch mutant drives (M1/M2 × REV/FINAL) with diff-verified mutants: the red drives the report names, re-driven on diff-verified copies (`cmp` against the `git show` pre-image pins byte-for-byte). Not built: the pin itself is the red test driving the M-class; the mutant batteries above carry the reds, no new member needed.

All three are session-side re-drives of the delivery's own claims; none rode an uncommitted synthetic member.
TRIPPED: none
PROPOSED ROWS: none
METHOD FRICTION: none — both documented forms (`scripts/check-tests --fast` and the module form) ran as documented, the review-acceptance scope gate ran first, and the review's scratch drives re-derived everything else; no interpreter, wrapper, or narrowed form was substituted anywhere.
INHERITED:
From the review brief `docs/research/phase6-janitor-window-review-brief.md` — its INHERITED block's still-open copy, taken verbatim: the exit-gate-runs chain's carried set as the closeout-status review's INHERITED presents it (the chain's carried blocks verbatim), then the exit-gate-runs find-blocks and that review's own carry line, as the brief moves them:
    F13 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried (the pin's stated candidate is not the one that flips; its next action is a later committed brief's text correction).

    F14 (P2, ORIGIN: DOCUMENT, carried) — the F8 red evidence is load-borne and the implementer report does not say so; its next action is the planner's correction of the carried note.

    F17 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried (the boundary-fix instruction's internal inconsistency; its next action is a later committed brief's text correction).

    F21 | SEVERITY: P2 | ORIGIN: DOCUMENT | STATE: carried (the report's battery line deviates from the brief's named command; its next action is a later step's or the planner's correction).

    F22 (P2, ORIGIN: BRIEF) — the truncated review-block copy of F17; a carried text-correction to the planner.

    F23 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried (the named red at REV is not reproducible; its next action is a later committed brief's text correction).

    F26 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried (the repoint's assertion form; its next action is a later committed brief's text correction).

    F28 | SEVERITY: P2 | ORIGIN: DOCUMENT | STATE: carried (the report's docs-baseline claim falsified; its next action is the planner's correction of the carried note).

    F29 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried (the probe-absorption vs anchor reading; its next action is a later committed brief's text correction).

    F30 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried (the named red is chimeric as literally written; its next action is a later committed brief's text correction and the cross-process janitor F15's own next action names).

    F31 | SEVERITY: P2 | ORIGIN: DOCUMENT | STATE: carried (the report's SHAPES field claims "none" while its own F30 block records a probed shape; its next action is a later step's or the planner's correction).

    F32 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried (the red clause's load regime; see FINDINGS above; its next action is a later committed brief's text correction, absorbed by F14's pending load-condition carry).

    F33 | SEVERITY: P2 | ORIGIN: DOCUMENT | STATE: carried (the report's docs-line file-count narrative; see FINDINGS above; its next action is the planner's correction of the carried note).

  From `docs/research/phase6-exit-gate-runs-review.md`, its find-blocks' received-carried items verbatim (the chain's F1/F2 are the exit-gate-runs implementer report's filings; F3/F4 its review's filings — all carried, none with an action a step of this chain has taken yet):
    F1 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried
    The brief's CONTEXT and RESIDUALS make the report commit the step's content commit and say `FINAL` names it, but the template also requires the report to paste the `check-scope` line, whose range carries `FINAL`'s sha — a sha that cannot be known before the commit that carries the report exists, and cannot be written into that commit's own tree (a fixed point). A one-commit realization is impossible.
    FIXED: no regression test — a committed brief's prose has none. Reading taken: the commit that first carries the report is `FINAL`; the report's two self-referential fields (`FINAL`'s value in the header and the `check-scope` VERIFIED line's range) are finalized in the immediately following commit, so `scripts/check-scope c418342 <FINAL> docs/research/phase6-exit-gate-runs-brief.md` reproduces the report's pasted line exactly against `FINAL`. Next action is the planner's: a later committed brief's text correction (rule 8's form) — this step delivers the reading and cannot correct the brief itself.
    F2 | SEVERITY: P2 | ORIGIN: BRIEF | STATE: carried
    The brief's step-3 store list names "the debug transcript" among the stores to scan, and `docs/record.md`'s campaign paragraph says the scan includes it, but the contract's "export exactly these" block omits `ORCHESTRATOR_DEBUG_TRANSCRIPT`, so the transcript (opt-in, default off; `resolve_transcript_path` returns `""` and nothing is written) does not exist for this run.
    FIXED: no regression test — a committed brief's text has none. Reading taken: the contract's exact export list governs, and the scan covers every store present (the transcript's absence is the contract's own omission, not a scan gap — the secret is proven absent from every file that exists). Next action is the planner's: align step 3's store list with the contract (or add the export if the transcript is wanted scanned).
    F3 | SEVERITY: P2 | ORIGIN: DOCUMENT | STATE: carried
    The implementer report's figure "The seven `budget_exhaustion=True` trials" miscounts its own enumeration (which names six trials: RECURSIVE_LAZY 1001/1002/1004, RECURSIVE_CONCURRENT 1001/1002/1004) and every countable surface of the run agrees on six: the narration census (`6 True / 10 False`), the runner's structured `trials` array (6), and exactly six `BUDGET_EXHAUSTION_RECORDED` rows across the trial stores (all 54 attempt-level `budget_exhausted` values are False — the "seventh" exists on no surface). The review brief's FOCUS carries the same "seven" twice.
    FIXED: no regression test — the figure lives in a committed report; the census is re-derived in this review's VERIFIED; next action the planner's (or a later step's) correction of the carried note, in the F28/F31/F33 form.
    F4 | SEVERITY: P2 | ORIGIN: CODE | STATE: carried
    The step's two commits under review (`67f5531`, `36e42a4`) carry no `Actor: IMPLEMENTER` trailer, against the method's mandate ("Every actor commits its own work with an `Actor: …` trailer, because … the trailer is the only thing that makes 'you must not be the author of the commits under review' checkable") and against the templates' commit instructions; the review brief's gate sentence asserts the trailer "carry" and its assertion is false here. The lapse is step-wide: since `c418342` five consecutive commits lack their actor trailer (`d1997b0`, `bc56b7c` — planner; `67f5531`, `36e42a4` — implementer; `715f425` — planner); the last trailerred commits are `c418342` (`Actor: OWNER`) and earlier. Reviewer authorship of the two commits was still verified independently (both were delivered from the campaign session's own workspace before this reviewer's session began; the commit author is the shared `gituser` identity, which the absent trailer would have disambiguated).
    FIXED: no regression test — a commit-message form the delivery cannot amend post-review without moving FINAL ("FINAL must not change after the report names it"); next action the planner's: restore the trailer discipline in practice (its corrections travel as rulings; a re-cut is not possible for a past commit) and let later steps carry it.

  Then each finding marked carried by this review, with its ID: F13, F14, F17, F21, F22, F23, F26, F28, F29, F30, F31, F32, F33 (received carried, kept carried — this review touched neither the brief-text defects they name nor any code path they cite); the exit-gate-runs chain's F1, F2 (received carried, kept carried — re-verified in the exit-gate-runs review's own VERIFIED and their premises unchanged here); its F3, F4 (received carried, kept carried — the trailer lapse its F4 filed recurs in this step's two commits as this report's own F1, and the six-vs-seven figure stands as re-derived); and this report's F1, F2, F3, F4 (filed by this review, carried — a commit-form lapse in the step's two commits, a battery-member scan-coupling in the janitor pin, a falsified hash cell in the exit-gate-runs Record row, and the brief's mandated "stores" noun — each with its named next action).

Then each finding marked carried by this review, with its ID: F13, F14, F17, F21, F22, F23, F26, F28, F29, F30, F31, F32, F33 (received carried, kept carried — this review touched neither the brief-text defects they name nor any code path they cite); the exit-gate-runs chain's F1, F2 (received carried, kept carried — both re-verified in the exit-gate-runs review's own VERIFIED, their premises unchanged here: this review's step touched neither the report fields they describe nor the store-scan contract they govern, and the delivered pin's re-scope touches only the battery member's window); and F3, F4 (received carried, kept carried — the six-vs-seven figure stands as re-derived by its own review and was closed actionless only by name, the substance waiting on the planner's carried-note correction; F4's premise extends: the planner's brief for this step, `e9a37db`, also carries no `Actor:` trailer — the sixth consecutive trailerless commit in the backlog F4 filed — while this step's own two commits under review (`a14afb2`, `819ddde`) carry their `Actor: IMPLEMENTER` trailer as that discipline asks).

The one fixing absorption this chain took: the closeout-status review's F2 — the janitor pin's whole-`gettempdir()` scan coupling, whose row cell named "a later brief's re-scope or serial-set move" — is fixed by this step: the later brief (`e9a37db`) took the re-scope reading and the implementer delivered it (its own F2, fixed with red evidence), re-driven red at REV and green at FINAL by this review (VERIFIED above). The closeout-status review's F1 (its step's trailer lapse), F3 (the exit-gate-runs REV-cell hash, corrected in the closeout write itself) and F4 (the "2326 stores" noun, corrected in the same write) remain closed actionless as the closeout row records them.

Received-set dispositions kept: F16 fixed by the report that filed it (the deadline-margin review's delivered margin, its red re-driven there); the closeout-status review's F2 fixed by this step (above); F11/F12/F15/F18/F19/F20/F24/F25/F27 stay fixed as recorded in the reports that filed them; F13/F14/F17/F21/F22/F23/F26/F28/F29/F30/F31/F32/F33 stay carried with their named owners and text corrections; the exit-gate-runs chain's F1–F4 stay carried with their named planner next actions.
