[Clear1nsight]

Problem and resulting behavior

A capture failure could be replaced by a later cleanup error. In the rejected v2 controller, a raw-stream cap failure followed by fsync failure could lose the cap cause, skip later descriptor closes, and prevent terminal accounting. A SPAWN_RETURNED ledger error could also leave acquired pipe handles outside the explicit cleanup set; final ledger close could override an earlier failure or apparent success.

The preserved v3 successor registers acquired pipes immediately, attempts each cleanup/sync/close independently, keeps the first failure primary and observed later failures secondary, quarantines failed ledger appends, and emits success only after accounting closes. This new isolated package publishes that source, its independent reviews, the rejected version and exact failure evidence, and reproducible fake-only tests. It establishes scoped source behavior and finite conformance; original C05 and full V03 remain open.

Contribution and source identity

New package only: research/h25-raw-pilot-failure-accounting, 68 files / 1,493,862 bytes, all Git mode 100644. There are 59 exact mapped retained files / 1,417,389 bytes. Manifest SHA256 e888f1a8436ed86a5c9b5d09d2c685cfafab59a83b95ba7e3fc1aecf7fa2397d; source-map SHA256 08daccd3bc4439ce93c0af63abb6a98ca058b2e41211877717dcd9f42bad557c.

Remote reviewed head edc2820ba023b804ef1351a88e2fe31bf8d489dd, tree 049436557a10ac13d7cc49147b8ececdce2c637d, sole parent/base 7c49fe32c55de02c2c873655c804b16e4c3fef0a. The local qualification commit is 761d5a58f58b7215d4c561c44288f081e1da2ff5, with the identical tree and parent. Git Data publication uses the connected account's author/committer; both commit identities and exact bytes are retained separately.

This is the directly assigned prerequisite control-family contribution authorized while original activity credit remains unchanged. No task picker, selected-frontier change, new kappa or performance claim is proposed. The controller prototype has frozen local context/authority requirements and is not a portable admitted native collector. The registered hook invokes only the static checker and fake-only harness; it never invokes controller main.

Validation completed locally

• Independent source review rejected v2 before execution. One bounded v3 repair and seven added combined-fault cases were independently reviewed. The resulting 24 unique methods passed once normally and once with -O; every process/proc/selector/group boundary is fake, with owned temporary data only.
• Exact publication/source review verified every package and origin byte, staged blob and mode. The static checker imports no controller or original module and launches no subprocess. Its seven archive-fault controls reject changed/missing/extra bytes and invalid scope/accounting metadata.
• One official registered replay passed metadata, the seven archive controls and the unchanged 24-method normal/optimized pair: PASS_REGISTERED_RESEARCH_PACKAGE ... outcome=partial.
• check_research_commit.py --ref 761d5a58f58b7215d4c561c44288f081e1da2ff5 --check-source-history --check-modes passed all 68 committed files. History coverage is declared-base ancestor availability, not complete scientific history. The unrelated missing base-PDF object in the local alternate cache remains unrepaired and explicitly unclaimed.
• The official selected-package export and --check-inputs passed. One package-only, read-only Git-free cold replay of the same three commands passed with empty executable PATH and fresh bytecode namespaces. All 68 exported bytes were compared again afterward; files remain 0444, directories 0555. Export SHA256 002708dda50dd26462d93bca3a69fe9e9468948e77221c87ff153411e4ca4540.
• Independent final execution/export cross-review passed at exact local commit/tree, checking 21 publication raw/time files, complete method names/exits and transport equality without another replay. Its retained receipt SHA256 is b919df06b2ff3a059aeddfa4bc3725a08b9697ebba0ce4e1ba45c8049a7e900a.

Reproduction from repository root, CPython 3.12 standard library only:

python3 -B scripts/research_package.py --package research/h25-raw-pilot-failure-accounting
python3 -B scripts/check_research_commit.py --package research/h25-raw-pilot-failure-accounting --check-source-history --check-modes

No timing threshold is used. The checker is reviewed for this exact small packet under watchdogs; its post-read size checks are not a general hostile-archive allocation guarantee.

Qualification and accounting limits

Real process cleanup, descendants, deadlines and a positive main queue remain untested. On partial raw-write failure, completed-block byte/hash counters may omit a retained file prefix; actual raw-file readback or an explicit unknown disposition is required. Delivered-byte capture is distinct from complete intended GCC emission, reporter fidelity, original-consumer equivalence, loader closure, C07 lifetime and complete C05.

The native eight-job pilot closed partial at source-plus-synthetic 2/4, with zero compiler queries, zero shell-stub calibrations and zero original baseline/V03 cases. Its proposed native reservation was never granted or spent. Original B1-B4 remain withheld. No original Root/provider/helper/native-TCB/entry, policy or namespace operation ran.

Original completion stays 43.5%, 22/49 activities; 2/10 full stage gates; 1/10 overall milestones; 24 components / 52 helpers. No criterion, denominator or weight changed. The app goal service is separately observed paused; documentation changes do not alter service status.

The preserved canonical policy record is 94.550533479 seconds. New source controls add 0.74 and local publication qualification adds 3.66, producing a supplemental 98.950533479 / 2400 policy seconds within the original initial 600. No new budget. GNU-time individual maxrss and policy debit are not a full physical invoice. Preparation, reviews, Pro, publication, final metadata and historical gaps remain OPEN_NOT_ZERO. The float-spelling metadata-readback failure is preserved; no tests were rerun to repair bookkeeping.

Related work and bounded coverage

Repository: https://github.com/CrocSwap/integer-mult-bounds . The exact 37-node known dependency graph retains its scope and historical metadata limits. PR6321 is the canonical technical handoff; PR6397 supplies the exact two-command C05 domain/criterion correction. PR6377 and PR6379 retain their distinct source-only/conditional results; PR6371, PR6373 and PR6374 retain separate entry/caller/stage constructors and unadmitted runtime obligations.

At 22:02:13 UTC, main was unchanged, the owned fork permissions were verified, and the first 30 open PRs by update were inspected. The plausible keyword matches PR6395 and PR6415 were checked by body and paths: they concern PR4440 integer Taylor evaluation and PR4345 matrix-header conjugation respectively, with separate sources/contracts/quantities. This is bounded overlap coverage; no global novelty, unpublished-work or all-current-source audit is claimed.

Open next questions

For the collector, first admit the actual command/context/funding and known-group/unknown-spawn termination disposition; retain partial writes and failed accounting honestly. This PR grants no native execution.

For C05, close the property-specific installed driver prefix/spec build frame and complete all named complements/currentness/fit. The archive records the Pro19 frontier; subsequent Pro20 work identifies /usr/libexec/gcc/ as a native-recipe candidate with a concrete gcc-cross/ competitor. Retained generator/parameter files are being analyzed for phase preservation; source-package parameters are not an actual build phase record. A complete final gcc.o export is one sufficient method, not a newly mandatory engineering criterion. Neither the candidate nor these tests qualifies the installed field or original C05.

After genuine source/configuration/parent/native/receiver/funding/C06/C07 admission, run the intact V01 and every original V03 case in both Python modes under unchanged bounds. No helper, prefix refusal or publication may replace those obligations.

Credit and merge scope

ChrisR/Clear1nsight retain contribution credit, alongside predecessor PR6029/PR5864/PR6321 work, Douglas R. Colkitt and upstream maintainers, and unpaidmoon/Cursor-Grok where retained source lineage records them. Substantial OpenAI Pro, Codex and GPT-6.1 Sol Max assistance is disclosed. Independent AI source/export reviews are scoped evidence, not independent human acceptance or a formal proof kernel. Exact retained notices/licenses and failed evidence are preserved. The unnecessary raw browser/sidebar snapshot was excluded from the public copy; its original remains local.

Door: two-way. Blast radius: isolated new opt-in research package. Shared production/source files and upstream tooling are unchanged. Hosted exact-head planning/replay is pending at submission and will be read back separately; local success is not hosted success.

[Clear1nsight]

Delivery checkpoint — reviewed archive and exact-head hosted replay

Verified head edc2820ba023b804ef1351a88e2fe31bf8d489dd, tree 049436557a10ac13d7cc49147b8ececdce2c637d, sole parent 7c49fe32c55de02c2c873655c804b16e4c3fef0a; all68 current changed-file blob IDs match the independently reviewed 1,493,862-byte package. Draft/open/unmerged, maintainer edits disabled.

Actual attempt1 hosted run41538532628 is completed SUCCESS: plan125596441658 and registered replay125598524758 both SUCCESS. No separate unchanged-inputs or original scientific job is credited.

Local registered replay, committed-byte/declared-base-ancestry/mode check, official selected export/input check and ONE read-only Git-free empty-PATH cold replay all passed. Seven archive-fault controls and24unique fake-only methods normal/O are reproducible; source, committed, exported, cold and transported68-file bytes/modes agree. Independent source and final export cross-reviews passed for this scope.

Canonical technical handoff · Coordination pointer. Both posts were read back. The completed contribution is preserved primary/secondary failure accounting and finite fake-only controls, with rejected source/one repair/bookkeeping failure retained. Real process/descendant/positive-main behavior remains untested; partial raw-write counters require actual-file readback or explicit unknown on failure.

Native pilot closed partial2/4; compiler queries0, shell calibrations0, originalbaseline/V03cases0, originalB1-B4withheld. Full C05/full V03 remain false; original43.5%,22/49activities,2/10stages,1/10milestones,24components52helpers unchanged, originalcredit0. Canonical94.550533479 remains; supplemental0.74+3.66 yields98.950533479policyseconds/2400(initial600), no new budget. All complete physical/outside/historical/publication/final-metadata costs stay OPEN_NOT_ZERO.

Current next research gate is an independently bound actual-build libexec-prefix property slice or qualified equivalent, retaining phase/flag/helper/context and complete original C05/C07/C06/intactV01/fullV03 obligations. Source availability, a matching literal, finite tests, documentation and publication do not supply those qualifications or change the separately paused goal service. No further local replay or automatic native/source-unit restart is authorized by this delivery checkpoint.