Repository navigation
feat(lifecycle): stop the current verdict's identities before proving completion - #100
Merged
Merged
Conversation
- Exercise counted cleanup, bounded replacement passes and preserved uncertain evidence. - Keep native no-signal and wrapper-custody assertions while requiring final completion evidence.
- Cross job failures, current identities and evidence availability without native host effects. - Replace interim manual-cleanup expectations while preserving custody and no-signal safeguards.
…tion - Preserve available cleanup through evidence and job failures, with fresh post-job and final verdicts. - Order represented children before ancestors and bound full-identity cleanup to two passes. - Normalize job-helper failures and keep completion proof separate from port admission.
- Describe bounded current-verdict cleanup, wrapper fencing and uncertainty refusals. - State ordering, native signaling, port and controller-adapter limits.
Keep all 630 reader-outcome cells and every cannot-say/manual refusal assertion. Use an explicit recovery assertion only when an irrelevant launcher leaves a counted runtime orphan.
Update only the counted-survivor detail expected from shared lifecycle admission. Keep both lock modes, the exact public refusal, one private error file and all cannot-say/census controls.
Compare forty complete-selection worlds with the pinned base ordering rule. Cover native lookup races, absent-run GPU and port results, unique refusals, failed-start context and platform-specific executable identity.
Recover the base server and wrapper subtree order from the verdict ancestry walk while retaining child precedence. Recheck ProcessLookupError, preserve absent no-op results, deduplicate refusals and report start failure before cleanup failure. Correct source ownership comments without changing the verdict or record parser.
State subtree ordering, interrupted-stop recovery and unique start/cleanup failures. Distinguish absent-record no-op results from attempted cleanup, including port waits and GPU fields. Describe accounted lookup-race handling without widening the native signaling guarantee.
Pin all four start-error/interrupt and cleanup-error/interrupt combinations by exact exception type and text. Exercise actual CLI routing without an OS signal and distinguish first-seen diagnostics from sorted uniqueness.
Catch ordinary cleanup exceptions only; preserve KeyboardInterrupt propagation through start and CLI. Keep combined start-first diagnostics when cleanup raises an error.
- Limit interrupted-stop persistence to a fenced writable record before final publication; retain the published result during the port wait. - Distinguish cleanup errors from interruption and preserve the original start exception after successful cleanup.
- Pin runtime-alias, live-foreign-parent, owned-parent, orphan, system-ancestor and runtime-subtree ownership boundaries. - Require both individually verified identities with reciprocal raw parent numbers to receive cleanup attempts without publishing completion.
- Leave aliased runtime processes and live outsiders' unrecorded subtrees untouched, visible to the final census and named for manual cleanup. - Preserve recorded cleanup while requiring a comparable earlier parent creation stamp for individually verified fallback edges.
- Account explicitly for the signal-eligibility census without allowing another status or admission classifier. - Bound the existing 3956-cell model assertion to its recorded, descendant and orphan worlds; its cells and expectations are unchanged.
- Describe eligible unrecorded leftovers, manual refusal for aliases and live outsiders' subtrees, and unchanged status/admission. - State the shared-runtime residual and the earlier-counter requirement for fallback ordering edges.
- Keep the real start admission refusal assertion while substituting only the native lock in the host-free Windows ownership cell. - Fix the revealed msvcrt fixture error; retain all twelve parent/platform cells.
- Cover both platforms and absent/readable run records with an unidentified outside parent and a runtime child subtree. - Keep the same absent-parent worlds as successful cleanup controls; pin live-set retention, manual refusal and proof revocation.
- Withhold unrecorded runtime candidates when the current outside parent number is enumerated, whether identified or unidentified. - Preserve the accounted reader's cleared-parent and platform-system exceptions.
- Say that an enumerated outside parent withholds runtime cleanup whether identified or not. - Disclose subreaper adoption and Windows unidentified parent-PID reuse as manual-cleanup limits.
- Cover runtime-only boundaries and changing parent evidence without native effects. - Require observed absence or replacement across all three completion paths; 49 red assertions at ef389f0.
- Remove executable-location eligibility and its cross-census ownership inference. - Retain every unrecorded identity observed by this stop and require reader-confirmed departure before any completion. - Keep recorded and verified-descendant ordering, two passes and the empty final census requirement.
- Replace runtime-only signal expectations with named refusal without changing status or admission. - Retain state/evidence product cardinalities and compare recorded signal order with the base in every state world. - Keep identity replacement and two-pass controls on verified descendants; preserve runtime-only counterworlds.
- Runtime-only processes are named and left for manual cleanup, as on the base, but cannot hide behind success. - Require observed departure on all completion paths and document shared-runtime interference. - Defer any wider signal authority to ownership established at launch.
- Pin exactly 320 state worlds where reviewed child precedence overrides the base recorded order. - Compare recorded and verified-descendant signals to the frozen reviewed order in every admitted state world. - Keep same-counter image changes as unknown identities and retain the P8 runtime-only survivor for manual cleanup.
- Add exact promotion and unavailable-departure diagnostics; one deleted-image identity is named or attempted once. - Pin nested recorded-wrapper labels and active return after a job-close callback removes the run record. - Red receipt has four intended failures and one corrected fixture accounting mismatch; no source change.
- Exclude attempted identities from the final left set, including later verified descendants. - Separate a final live observation from an unproved historical departure in operator advice. - Reuse the process reader executable equivalence for attempted and remembered identity keys.
- Keep all 45 departure cells and completion assertions; twelve uncertain-history cells require cautious identity advice. - Current final-live cells keep the manual-cleanup text.
- State current-role selection and stop-local uncertainty without promising persistent verdict memory; link #102. - Separate current-live manual cleanup from cautious unproved-departure advice. - Describe the tested job-close record-loss return, not active cleanup with a record absent throughout.
- Keep all 300 job/identity/availability cells and proof assertions. - Seven final-unavailable location-only cells require historical observation advice, not a claim of current liveness.
- Cover final-only promotion and attempted-descendant demotion on both platform flags, with exact text and retained refusal/proof assertions. - Red at 71b93e7: two promotion assertion failures, zero errors; both reverse controls pass.
- Partition final-live survivors by current role and attempted identity; historical memory is consulted only outside the final live set. - Keep unattempted location-only manual advice distinct from generic survivors and uncertain departures.
- Keep one identity and one attempt across deleted-image observations; show the final image spelling for the surviving unrecorded process. - Windows reports the final case spelling once; POSIX retains the earlier distinct-case identity only as uncertain history.
…ntities - Describe final-only promotion as a survivor under its descendant role, eligible for a repeated stop. - Reserve earlier-observation memory for identities absent from the final live set.
- Do not promise an attempt for a role first established by the final verdict after both passes are over.
- Append independent coarse-stamp reissue worlds with and without a token; both reproduce stopped while the replacement is alive. - Token red receipt shows tree_verified_empty published; token fixture explicitly enters Windows lifecycle custody.
- Drop earlier location-only memory at the attempt, not by membership at completion. - Later observations of an indistinguishable key are remembered anew and require proved departure.
- Append all three-verdict observation sequences with applicable departure outcomes across absent, ordinary and proof-bearing returns. - Pin 128 reachable worlds and 34 explicitly unreachable cells; use a plain event-order oracle independent of stop memory sets.
- Append repeated stop from the exact first-call state, supplied-job absent-record authority, and final-port interrupt publication tests. - All three ties passed against the predecessor; the token interrupt fixture explicitly uses Windows lifecycle custody.
- An attempt supersedes earlier memory; an observation after the attempt is remembered anew. - Final departure checks cannot be bypassed by membership in the attempted set.
- Cover all four job observations and all three ordinary observations; pin 236 reachable worlds and 34 impossible absent-record cells. - Keep identity-once assertions on manual and historical lines; survivor expectations remain role-by-role. - The independent reader oracle has not yet been opened.
- Restrict identity-once wording to manual and historical refusal lines; final survivors retain every recorded role.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stop owned lifecycle identities before proving completion
Invariant
Where structural custody permits it, lifecycle stop attempts only recorded identities and their verified descendants in at most two passes, and reports stopped or records an empty-tree proof only after a readable unfiltered empty final census, observed absence or replacement of every runtime-only identity it left, and no record, census or job failure.
Mechanism
stop_incompleteand its prior empty-tree proof revoked before native effects. An absent or unreadable run is never created or overwritten.ProcessLookupErroris followed by an accounted recheck and cleanup continues; only proved absence or reuse dismisses failure, never individual no-match or cannot-say.wrapperwhere the base reportedserver-descendant; the identity sequence follows the base preference, subject to child-before-ancestor precedence.What changes for an operator
Lifecycle stop and start's generation-guarded failure cleanup use the same loop. Identical diagnostic texts are reported once, in first-seen order. If start and cleanup both fail with errors, the lifecycle error names the start failure first and cleanup failure second. If start was interrupted and cleanup fails with an error, that lifecycle error still names both in the same order. An interrupt during cleanup remains an interrupt, as before this change; successful cleanup re-raises the original start exception. CLI restart cannot proceed through a refused stop. G16 and G20 in
scripts/realhost_gates.pyinherit the lifecycle result; their other evidence rules are unchanged, as are the indirect restart/cleanup calls inscripts/requalify.py.Finite models
These are overlapping bounded domains, not a sum of independent native situations. Product counts describe only their enumerated inputs; qualification scope and reproduction commands are below.
The state/evidence products include 216 of 3,956 state cells and 5,712 of 16,128 evidence cells with runtime-only signal refusal; their status/admission attribution is unchanged. The state model pins exactly 320 base-order exceptions: recorded children precede recorded ancestors; all other recorded order is the base order. The forty complete-base-selection worlds retain their full sequence comparison; the state model permits additional signals over the base only to verified descendants. Fixture-driven replacement observations and separate verified-descendant controls keep the same boundaries observable without assuming a forbidden runtime-only signal.
Limits
ProcessLookupErrornow use the accounted recheck and allow remaining cleanup. Other native errors propagate and may interrupt cleanup before a final census; there is no atomic signal claim.owned_gpus_released: null. A record read earlier can disappear during cleanup: retained identities or a job token can still drive cleanup, while a later verdict observing absence disables further writes. Active calls take the normal return and final port wait, with attempted PIDs plus retained engine/server-tree PIDs for GPU ownership and a boolean release result. The host-free test removes an initially present record in the fake job-close callback after job cleanup and checks that active return without recreation; it does not claim active cleanup with the record absent throughout or that native close removes the file.statusandstartdo. Except for that successful absent-record no-op, the bounded port wait follows cleanup and any permitted record write, including absent-record failure paths. Errors from that final post-cleanup port wait propagate; where publication is permitted, the final record write precedes the wait. Absent or unreadable records are not published.stop_incompleteand running stop again is the recovery. An interruption during the port wait after publication leaves the published result.HostAdapter.stopin the experiment loop remains separate: it writes the sametree_verified_emptytoken after retained native-job termination and token/observed-identity enumeration, without this loop's final runtime census. This is a narrower proof denominator, not a claim that every writer of that field was migrated.Base dependency and reproducible qualification
The base is
ccad158d6b7204f91e221ac48d6d4c2a2933da58. Its verdict reader refuses a repeated JSON key at any depth. Stop consumes that unreadable-record verdict; this patch does not implement a second parser rule.Author-run qualification at
e6b92e2ccb82e8b7e2a1aece770a99dbd336cf62:OK (skipped=10, expected failures=4), 503.790 seconds; 4,028 owned subprocesses, none unreaped or left. Skip and expected-failure policies are unchanged.These measurements are author-run host-free evidence, not native-host qualification. The repository commands below reproduce the focused/model selections, whole suite and checks; eight selected observation/identity controls are self-contained below.
Run from the repository root with Python 3.13 and the repository's development dependencies. The focused selection is:
For the ownership/ordering products and their individual counts:
The eight controls below exercise final-role reporting, diagnostic classes, executable equivalence and all three temporal-memory mistakes. Their baseline contains seven checked-in methods; they do not reproduce the whole campaign:
CLI cleanup routing is tested without sending an OS signal: cleanup interruption escapes as
KeyboardInterruptinstead of returning ordinary-error status 4; no numeric signal-termination status is claimed by that test.The full suite uses an existing materialized development environment, without fetching or installing dependencies.
Set
OMP_STRATA_OMP_BINARY,OMP_STRATA_STRATA_SRC,OMP_STRATA_STRATA_PYTHONandOMP_STRATA_PROFILEto thatenvironment's executable, runtime source, Python and profile paths, respectively:
PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests -t .Repository and documentation checks, including every checked-in profile, route and manifest: