F07: strengthen cancellation cleanup and output-byte regression assertions

@De-pitcher

The original I01/I02 fixes and detecting regressions are independently accepted locally; full F07 acceptance remains pending.

Scope: explicitly apply test policy; require propagated CancelledError, committed bytes, unsealed conservative undo and owned-child teardown; cover repeated cancellation/noncancelled sealing failure. Add exact JSON UTF-8 output byte assertions and independent 62/63/64 boundaries for Unicode/escaped content. Preserve existing regression evidence.

Accountable GitHub owner: @De-pitcher. Implementation worker: Antigravity (replacing Nemotron 3 Ultra for #76). Independent reviewer: Codex. Do not duplicate that worker.

Acceptance: focused source-bound red/green evidence, current whole-tree lint/format/types, fresh correction manifest. No live/remote/release acceptance implied.

Independent review of the latest 13-test submission: 13 passed, no skips; actual ruff check ., ruff format --check . and mypy . pass (310/181 files). Original I01/I02 implementation acceptance stands. New policy configuration, committed-byte check, noncancelled failure case and Unicode/escape byte assertions are useful improvements.

#76 remains open. Its stated acceptance still requires:

Require propagated CancelledError; fail on ANY returned result, including an ordinary failure whose text says cancelled. Assert the original request is pending before controlled cancellation.
Assert the retained checkpoint is unsealed, undo refuses unknown post-state without rewriting committed bytes, and observed mutation/sealing helper PIDs are gone. Existing unsealed lines are comments after an early return, and the marker PID is never checked. Cancel/drain pending requests and close owned resources in failure paths.
Deliver repeated cancellation to the SAME request while its post-state cleanup is pending. The current test instead awaits task one and immediately cancels a newly created task two. Use distinct fresh phase markers and verify the same journal/teardown facts. Also check those facts for noncancelled sealing failure.

Keep the existing Nemotron worker; do not duplicate it. Next action is these focused assertions under #76, then a scoped issue-linked PR with source-bound raw evidence. Preserve dirty F07 source and previous test/evidence bytes; no blanket staging or unrelated fixes. #77-#79 remain separate pending work.

Binding: historical 172 entries still 556ed8d017ef7d6bfa9ec7b61db6c9790e940adfba4a601d212261e7ceaedf0c; new test SHA256 208a415bbfda6de77090c0a8c8e7fcf1e532be54dc7601da4c8f7032f3098b95. Explicit 173-entry union 928816d90d5942b1dab0571bb1dd6a6c8bfd2a931a878a9fb20d6edfa9bdecb8. Historical 172 identity alone does not bind the new test.

Local reviewer evidence, not yet published in a PR: docs/audits/f07-issue76-review-2026-10-12/REVIEW.md, exact source snapshot, manifest and complete command JSON logs. Independent focused count is 13; reported 86-test total remains worker evidence. No full F07/production/installed/external acceptance granted.

Round-two independent review: 13 passed, no skips, 14.92s. Actual whole-tree lint/format/types pass (312/181 files). Strict propagated CancelledError requirement is now satisfied. Original I01/I02 fixes and improved I02 checks remain accepted locally.

Two existing requirements still keep #76 open:

Second cancellation must interrupt pending cleanup. The same-task test awaits the first cancellation to completion before cancelling again. The second await only observes the completed task's stored CancelledError. Wait for a separate fresh post-state-helper marker, assert the original task is pending, cancel that same request again, then await it once to require CancelledError. Use bounded readiness and distinct mutation/cleanup markers. Both cancel calls should request cancellation on a pending task.
Assert journal and process facts explicitly. The unsealed-state block is still comments, not assertions. Committed bytes do not prove unsealed state. Read/assert retained snapshot.sealed is false, verify rollback refuses unknown post-state without changing committed bytes, and read/check mutation/sealing marker PIDs are no longer alive. Avoid early return before these checks; cancel/drain pending requests and close owned resources in failure paths. Include these checks for noncancelled sealing failure.

These are the original issue acceptance criteria, not additional scope. Keep the existing worker; don't duplicate it. No open PR was found. When the focused requirements pass, submit a scoped Closes #76 PR and raw source-bound evidence, preserving the dirty F07 foundation and old test bytes. #77-#79 remain separate pending work.

Reviewed new test SHA256: f106375efdf673cdd1778f13235ad9c248c21334ab56a5ff791c516cdf60bc13. Historical172 identity remains 556ed8d017ef7d6bfa9ec7b61db6c9790e940adfba4a601d212261e7ceaedf0c; explicit173 union is 75d23eddbdd2ea2a2ff073029cbad947f6b45516ac6a9ab8eb79d4a9d8defa8a. Local evidence (not yet published): docs/audits/f07-issue76-round2-review-2026-10-12/REVIEW.md, snapshot, manifest and full command JSON logs. The claimed86 total remains worker evidence; no full acceptance granted.

Round-three review: 13 tests pass, no skips,14.05s; whole-tree lint/format/types pass(314/181). New unsealed-checkpoint and undo-refusal assertions execute and pass. Strict cancellation and I02 checks remain accepted.

The permanent repeated test still awaits completion before subsequent cancel calls; duplicating the block and writing STILL PENDING in comments does not test pending cleanup. It has no native PID-liveness assertion. Cancellation propagation alone does not prove child teardown.

I independently verified actual production behavior with a concrete working fixture: separate actual atomic-commit and parent post-state-helper PID markers, original request pending, two cancel requests both returning true, CancelledError, unsealed checkpoint, committed bytes unchanged, and BOTH owned helpers gone/exited at request completion and after context closure. This controlled Windows probe passes; no production edits. MutationPID24456, sealingPID548,0.122s after first cancel.

Use this reviewer fixture as the exact permanent-test design, rather than another comment-only revision:

docs/audits/f07-issue76-round3-review-2026-10-12/repeated-cancellation-owned-probe.py.txt
Raw result: repeated-cancellation-owned-probe.json in that directory.

Required order: observe atomic commit → first cancel → observe DISTINCT parent-cleanup readiness → assert SAME original task pending → second cancel before await → require CancelledError → assert unsealed checkpoint, refused undo/no reverted files/committed bytes preserved, and observed helper PIDs gone. Cancel/drain outstanding requests and close owned resources in failure paths. Apply journal/ownership facts to main and noncancelled failure cases. These are existing #76 requirements; production behavior is now independently verified, permanent checks still need completion.

Keep existing worker ownership. No open PR was found; submit scoped Closes #76 PR when the permanent checks pass. Preserve dirty foundation and prior test bytes. #77-#79 remain separate; no full F07/artifact/external acceptance.

Exact submission: test SHA5052e9deb214a4910d4be00273b603e527f6533f35380e4d0534bb80917762ff; historical172 identity556ed8d017ef7d6bfa9ec7b61db6c9790e940adfba4a601d212261e7ceaedf0c; explicit173 uniona7f48fcfb48e2d6fcedb6cf05eddcae4de322b43c0d6e4ba72c221b62557b680. Local full review/logs/snapshot at docs/audits/f07-issue76-round3-review-2026-10-12/REVIEW.md (not yet published in a PR).

Round-four independent result: 13 passed, no skips,14.71s; actual whole-tree lint/format/types pass(316/181). Native PID checks now execute and pass for the recorded PID. Strict cancellation and journal/undo checks remain accepted. Production behavior was already independently verified in the strict reviewer fixture.

One cleanup-phase fixture correction remains: two consecutive cancel calls do not let the event loop process cancellation number one before number two. The fingerprint marker already existed before cancellation; it does not identify the new parent post-state cleanup helper.

Use this order, with the exact helper hooks/bounded readiness from the already-working reviewer probe:

observe atomic commit and capture mutation PID
assert pending.cancel()
await fresh DISTINCT parent-cleanup marker (bounded wait)
capture cleanup PID; assert not pending.done()
assert pending.cancel()
require CancelledError on final await
assert unsealed/undo refusal/committed bytes preserved
assert BOTH captured PIDs are no longer live

Fixture: docs/audits/f07-issue76-round3-review-2026-10-12/repeated-cancellation-owned-probe.py.txt; its saved result proves this sequence works on current production code. Convert that sequence directly; don't substitute immediate double cancellation or fixed sleep. Keep current PID checks, but use separate phase markers so one overwritten marker cannot lose the original PID. Close Windows query handles in finally; cancel/drain pending fixture tasks and close owned resources on failure.

#76 stays open. Existing worker owns this change; no duplicate worker. No open PR found. When ready, submit source-bound scoped Closes #76 PR; document prerequisite dirty-baseline integration rather than blanket staging it. #77-#79 stay separate.

Current test SHAe4b38550e123442d85461c744fbded6a6be31e817d7bc3eae361b6eee624801c; historical172 remains556ed8d017ef7d6bfa9ec7b61db6c9790e940adfba4a601d212261e7ceaedf0c; explicit173 union1c3dcd62bf3ce4b544acc4385232702a95b697552eeeb4a8e3fb8030497ba6f0. Local report/logs/source snapshot: docs/audits/f07-issue76-round4-review-2026-10-12/REVIEW.md (not yet published). No fresh86/full-suite or release acceptance claimed.

Round-five submission rejected: new production policy regression. Independent13 tests pass/no skips and actual lint/format/types pass(318/181), but a focused owned-helper probe reproduces an ungranted native write.

R01 high — remove production marker instrumentation. _io_worker.py now writes request.cleanup_marker directly, before write-capability/workspace validation. An actual fingerprint query with allow_write=false overwrote a pre-existing sibling file OUTSIDE the granted workspace, despite direct policy.path(write=True) denying it. Fingerprint returned successfully. All files were reviewer-owned under repository temp; no user file targeted. Local raw proof: docs/audits/f07-issue76-round5-review-2026-10-12/marker-policy-probe.json and its script.

This hook is unnecessary production behavior. Preserve submitted bytes, remove only your introduced marker/import changes to restore the prior fixed operation, and keep controlled marker hooks inside the test-generated helper shim. Do NOT broaden policy roots or enable writes to hide the failure. Original fingerprint import was already present; the claimed import fix is not reflected in the actual delta.

R02 — cleanup marker is still premature. slow_fingerprint writes fingerprint.pid AND cleanup_ready.pid together BEFORE first cancellation. Since the first marker was already awaited, the cleanup marker also already exists; the subsequent wait returns immediately. It proves neither new parent cleanup nor pending second-cancel timing. Payload.cleanup_marker does not by itself become the parent fingerprint request's top-level field.

Use the EXACT passing fixture already supplied: docs/audits/f07-issue76-round3-review-2026-10-12/repeated-cancellation-owned-probe.py.txt. Its test shim uses actual os.replace commit signaling, and execute(request) writes a DISTINCT marker ONLY for the parent operation=fingerprint. No production modifications needed. First cancel → bounded fresh cleanup readiness → assert original pending → second cancel returns true → await CancelledError → unsealed/undo/bytes facts → BOTH captured PIDs non-live. Preserve failure-path cancel/drain/resource closure.

Source binding corrected independently: current172 identitybdf2dd4eae85850d428a206bb71df7eecaf71668fa02f23c7203e098e757a898, THREE production paths changed. New helper SHA6cfd8589c08713c8a46486260215832a17c511291ee92910e52b3b32af61ff72; test SHA084e28d79011d9b6dcdfe4dad6eeb0aab6d33523eb1646d887150126a20f82eb; explicit173 uniond555e80b70275d33a1723ecc1930ec601e8661d89e71919d579474e7357f5981. The reported556ed8d0/two-file scope is stale.

#76 remains open/in-progress. Existing worker owns correction; no duplicate worker or reviewer production edits. Restore boundary first, then permanent fixture and scoped Closes #76 PR. No open PR found; describe prerequisite baseline integration without blanket staging. #77-#79 remain pending. Full local report/raw logs/snapshots: docs/audits/f07-issue76-round5-review-2026-10-12/REVIEW.md (not yet published). Historical pre-hook I01/I02 acceptance stands at its own source; current hook is not accepted.

Production marker regression closed locally. Repeated actual read-only fingerprint probe now leaves the outside-granted-workspace file byte-for-byte unchanged. No production marker hook remains. Independent13 tests/no skips pass14.29s; whole-tree lint/format/types pass(320/181).

Source claim still incorrect: current172 raw-byte identity is8b32e71da7505b535417ecb61f0ccbc9c0a17a9625c426f08895d0b97cf64b7f, with THREE changed production paths. _io_worker.py still differs by Path import placement/raise indentation; its SHA is2bf1fdbf7b5d83db025ba650d47000bfc46c582457dc72e926f6d3b22784a630. Marker removal fixes behavior, not the historical byte hash.

Permanent cleanup-phase requirement remains: test file SHAe4b38550e123442d85461c744fbded6a6be31e817d7bc3eae361b6eee624801c is EXACTLY round-four's submission. Two cancel calls are back-to-back; the fingerprint marker already exists before cancellation. This is duplicate requests before cancellation delivery, not interruption after parent cleanup starts.

Please convert the already-passing reviewer fixture DIRECTLY: docs/audits/f07-issue76-round3-review-2026-10-12/repeated-cancellation-owned-probe.py.txt. Keep its markers in the generated test shim ONLY. os.replace signals actual commit; execute(request) signals only the distinct parent operation=fingerprint. First cancel → bounded wait for NEW cleanup marker → original request asserted pending → second cancel BEFORE await → CancelledError → journal/undo/unchanged-byte assertions and BOTH captured PIDs gone. Do not substitute the already-existing fingerprint marker or add production hooks. Include cancel/drain/resource and Windows handle cleanup in failure paths.

#76 remains open/in-progress; existing worker ownership unchanged. Next submission should include the corrected fixture in a scoped Closes #76 PR with accurate fresh source binding; no open PR currently found. Preserve the dirty foundation and explain prerequisite integration rather than blanket staging. #77-#79 remain separate.

Local report, probe and raw logs: docs/audits/f07-issue76-round6-review-2026-10-12/REVIEW.md (not yet published). Explicit173 union403f79149b4406abddf6d531c082af60088e9426aed8df9c77eac4584f064515. Prior original-fix acceptance stays historical; no full F07/artifact/external/live/release acceptance granted.

Antigravity handoff independently reviewed. The required cleanup-phase sequence is now implemented correctly: actual commit → first cancel → distinct parent-cleanup marker → same original task pending → second cancel accepted → final CancelledError → unsealed checkpoint/undo refusal/committed bytes preserved. This requirement is accepted locally. Worker role in #76 updated to Antigravity; accountable assignee remains De-pitcher.

Independent 13 passed/no skips,9.40s. Actual whole-tree lint/format/types pass(322/181). Reported172/173 identities and24591-byte test hash match fresh bytes exactly.

Two small original-requirement assertions remain before local test acceptance:

check_alive must fail on native query errors other than87. It currently returns false for ANY OpenProcess failure. A reviewer negative control using the exact function proves error5(access denied) is incorrectly accepted as a dead process. Only87 proves absence; otherwise raise ctypes.WinError(last_err) or fail an assertion. Add a negative query-error control. Keep the working handle/exit-code/finally logic.
Add assert rollback_result['reverted_files'] == []; the handoff claims this verification but the code omits that field assertion.

No phase redesign or production modification is requested. Preserve prior bytes and return fresh source-bound focused/raw evidence for those two edits. Independent local records: docs/audits/f07-issue76-antigravity-review-2026-10-12/REVIEW.md, liveness-query-negative-control.json, raw test/static logs and source snapshots (not yet published in a PR).

PR prerequisite: main lacks bounds.py, policy.py, runtime.py, io_boundary.py, processes.py and storage_limits.py. Thus the four-file PR proposal cannot stand alone; these are actual runtime dependencies. Planning-only #83 now records the required exact dependency/integration proposal. It authorizes no foundation publication, no extra execution worker and no work by Antigravity outside #76. Do not blanket-stage unrelated dirty source or silently enlarge the PR.

#76 remains open for these assertions and integrated contribution. #77-#79 remain separate. Binding:1728b32e71da7505b535417ecb61f0ccbc9c0a17a9625c426f08895d0b97cf64b7f; teste450f83cf0439fc5b5227d69c5110e9da43443b5b4572506e25aee2a82271a63;1730593b97f1596c5863a62ba072fc22d29bb0e627a604a99edcdf850882e5073d5. No production/external/artifact/release acceptance expanded.

Both requested code corrections are now independently verified: actual check_alive fails closed for errors5/1, and rollback explicitly asserts reverted_files==[]. Cleanup sequencing remains correct. 14 tests pass/no skips,9.41s; actual lint/format/types pass(324/181). Submitted hashes and172/173 identities match fresh bytes exactly. No production changes observed.

One negative-control wiring correction remains: the new test defines and calls its own query_alive copy, while the cancellation test uses a separate check_alive. It must exercise the SAME function to protect the actual liveness assertion.

I verified the actual helper directly:87→false,5→PermissionError,1→OSError. Then, in a disposable in-memory AST/module ONLY, I regressed the actual check_alive to return false on error5 and ran the submitted negative control unchanged. The negative control STILL PASSED because it called query_alive instead. No worker/source file mutation and no actual process targeted. Raw proof: docs/audits/f07-issue76-antigravity-corrections-review-2026-10-12/actual-liveness-control.json and its verification script.

Exact next edit: extract ONE shared test PID helper, call it from both the cancellation test and negative control, and remove the copied query implementation. Keep87/5/1 cases, correct native handles, accepted phase sequence and rollback checks. Verify the negative control rejects a fail-open mutation of that same helper. This completes the original meaningful negative-control requirement; no new production behavior or fixture redesign is requested.

Antigravity remains #76-only. Preserve prior bytes, return raw focused/gate evidence and updated source binding. #76 remains open; PR integration awaits the separate #83 dependency plan. #77-#79 unchanged, no blanket staging or publication. Local full report/logs/snapshot: docs/audits/f07-issue76-antigravity-corrections-review-2026-10-12/REVIEW.md (not yet published).

Binding:test401d72f978398e76e49a3c6ba1d8d726508f71ae10bcc72a606e4f0ab5b06faf;1728b32e71da7505b535417ecb61f0ccbc9c0a17a9625c426f08895d0b97cf64b7f;173ece37d0f1cd53c7fe28364731ad5702f714513fa85bbdbc6d1f7c73c78b8b816. Both behavior corrections accepted locally; remaining issue is permanent control linkage, not another functional failure.

Independent supervisor review: local #76 regression implementation ACCEPTED.

The single shared check_alive now serves both real PID-teardown assertions and the Windows negative control. Independently changing the actual helper to fail open on access denial makes the permanent control reject it (DID NOT RAISE PermissionError). Error87/5/1 branches verified. Pending-cleanup second cancellation, CancelledError propagation, unsealed checkpoints, rollback refusal with reverted_files == [], preserved bytes and both helper teardown assertions pass.

Independent sequential checks: 14 passed in 9.72s; ruff check . passed; ruff format --check . 326 files; mypy . 181 files. Test SHA-256 0266b7d4ca0827fc2b205276c4902da12365bf42e12304dab4aafe12091dbab2; historical172 8b32e71da7505b535417ecb61f0ccbc9c0a17a9625c426f08895d0b97cf64b7f; expanded173 e0abd9c02f641daee2850aaa31687922b630ae05e05805b4feda8913b40e3761. Raw commands/logs, source snapshot and mutation evidence: docs/audits/f07-issue76-antigravity-shared-helper-review-2026-10-12/.

Issue stays OPEN and integration-blocked on #83. This accepts local regression work, not a standalone PR or release. Antigravity has no further #76 corrections; do not expand into #77–#79 or unpublished foundation. Next authorized supervisor action is review of the planning-only #83 exact prerequisite/dependency inventory and issue-linked integration sequence. No blanket staging or foundation publication. Remaining release/external/live gates unchanged.

Issue83 prerequisite planning independently accepted; implementation tracked under assigned issue #84. #76 local regression acceptance unchanged. Integration remains blocked until the narrow foundation is implemented, independently reviewed and integrated through required PR checks. No new #76 implementation requested yet.

Navigation Menu

Repository navigation

F07: strengthen cancellation cleanup and output-byte regression assertions #76

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

F07: strengthen cancellation cleanup and output-byte regression assertions #76

Description

Activity

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 12, 2026

De-pitcher commented on Oct 13, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions