Quake II custody retirement service repair, 0844

The production custody provider now batches irreversible EXIT/terminal-wait records in an ordered buffer capped at 1 MiB and 4096 records. Saturation flushes synchronously before accepting another record; an oversized individual record bypasses the buffer after flushing its prefix. There is no worker, thread, unbounded queue, dropped record, synthetic wait or grace increase.

Registration, creation/lineage and EXEC admission retain synchronous durability. Every queued executable CONT explicitly flushes the ordered prefix first. Only a peeked real terminal status or kernel EXIT stop selects deferred service; the consumed status still controls retirement, GETEVENTMSG, EXEC transfer and resumption. EXIT CONT can finish irreversible death without per-record disk service. Every terminal receipt still comes from actual waitpid after WNOWAIT. Existing fork/CONT selection fairness, identity checks, nonleader EXEC binding, kernel lineage requirements and separate command EOF checks are retained.

Final close flushes all pending evidence, writes the gzip footer and fsyncs. Success still requires that close and command-log durability succeed. Journal write/fsync errors latch permanently; an error while logging failure cannot prevent identity-checked containment. Initial registration now sits inside that containment scope. Abrupt death can leave a durable prefix with an unwritten retirement suffix: missing final durability/drain remains explicitly incomplete, as tested by the existing hard-kill controls. This does not claim durable per-retirement records before each wait. Admission durability remains required. Bounded saturation/storage stalls may still block; arbitrary storage failure does not acquire a successful quit classification.

Evidence: scratch/runs/20261109-quake2-custody-retirement-repair-0844/. composed-final/ uses the actual composed Quake II driver, Python Caller, numbered transport, close helper and production ptrace supervisor with an inert 110-thread Python browser substitute. The old provider is recovered directly from Git blob c604ccf084212eaeeff2d0fd92c65fbd535360ab, without copying sources. At identical test-only 50 ms journal-service delay after ordinary close starts:

Provider Service writes after close Close elapsed Result
Old production 270 11767 ms TERM, exit0, EOF, normal:false; quit rejected
Repaired production 3 72 ms No signals, exit0, EOF, normal:true; numbered quit ACK

Both controls record 122 owned / 122 actual reaps, complete custody and command EOF. The repaired unpaced control also accepts ordinary quit. Journals have contiguous sequence numbers, valid gzip CRC/footer and matching original byte counts/SHA256. Supervisor, caller children and independent prearmed watchdogs are actually waited with separate EOF; watchdogs do not fire. No guest CDP occurs after Browser.close. Runtime/ACK/deadline caps, signal-normal classification and staged-root/startup gates are unchanged.

Validation: 46 custody tests pass, including byte/record saturation, oversized record ordering, durable executable-CONT prefixes, admission/saturation/final service failures, footer-sync refusal, cancellation, fork lineage, EXIT/CONT fairness, EXEC transfer/rejection, adoption and missing waits/EOF. All 21 actual caller composition scenarios and 10 close-helper tests pass, retaining TERM, KILL, ACK-stall, inherited EOF delay and cancellation rejection controls. flock /tmp/wine-build.lock bash tools/build.sh passes; CPU remains 2185596 bytes, SHA256 a320d4c85156644fd12b6134d66b11784f3100b9b135756206c087cdacd43295. Initial test failures exposed disk-only scheduling probes and a fixture file publication race; their logs are retained and excluded from final acceptance. Probes now observe recorded EXIT events without forcing service, while checking durable registration and admission separately.

Historical 0820 ACK-to-TERM still has 393 records, 94 EXIT stops and 93 actual reaps. Historical per-syscall latency remains UNKNOWN; this controlled service dependency repair does not establish an exclusive historical fsync cause. No Chrome, guest, original, box, agent/worker, deployment, host.js, WAT, HeroesII or Claude work occurred. Git base plus small owned patch and gzip logs provide recovery; source/media/module archives were not made. Existing source-matched proofs are referenced by Git rather than copied. Free /home remained above 1.5 GB. Box ID/state/expiry/stop: NULL / NOT_CREATED / NULL / NULL.

Next: a separately authorized fresh integration/native run with current provider pins; this offline result makes no gameplay or native qualification claim. Single-task STOP after publication.