[joelteply]

A slot waiter retained its endpoint read guard while releasing and later reacquiring its adapter semaphore. A later admission could hold that semaphore behind a queued lifecycle writer, completing a lock cycle. Release the entire admission before waiting on the registered slot-release notification, then reacquire in the existing permit-to-reader order and resolve the current generation pool.

Keep two distinct budgets: only endpoint-readiness waits spend cumulative readiness patience; the existing absolute request deadline bounds the whole admission, including permit and slot waits. Both streaming entrypoints pass these distinct budgets instead of pre-minimizing them.

Five regressions cover the lock cycle, transition draining running turns without draining the adapter queue, replacement-generation pool identity after a 300-second slot wait and a 20-second closed endpoint, readiness patience plus cancellation cleanup, and request deadlines during semaphore and all-pinned slot waits. Explicit cumulative debit across multiple successful outer retries is not separately exercised.

Validation: independent source approval at e992346; current 7adeb78 adds only the review-requested deadline comment correction. Fresh exact-head CI is required. No local compilation beside serving. Earlier 56c83d3 CI passed all four then-existing engine regressions (engine 1150 passed, kernel 1230 passed, core 6188 passed); that does not validate the new budget semantics. Merged #5190 (4b82aa8) is inherited at 01a6a1a with independent delta source approval. Current 014f41b only normalizes indentation after that approval (git diff -w empty). Fresh final-head CI is required. Pre-inheritance 7adeb78 CI also passed, validating the new readiness-budget regressions before the telemetry merge.

Baseline limitation: regression-only 7d1ba97 run 39890937094 failed in the old kernel disk-eviction fixture before engine tests ran. This is not an executed red lock-cycle proof. The lock cycle is independently source-traced; Intel live wedge causation remains unproven and needs owner acceptance after reviewed integration.

[joelteply]

Exact-head CI attempt1 at014f41b6a hit the job45-minute limit, confirmed by GitHub check annotation. This is not a test failure: core6004, kernel1230, engine1152 and hands192 tests passed; four installed-release CLI cases passed; doc guards logged2 passed as cancellation arrived. Database and binding-drift steps did not run, so the overall gate remains incomplete.

The measured cause was Azure apt mirror download throughput: dependency setup15m21 versus13s in the previous passing run; cmake14MB took a226.6s log gap and libclang30MB a526.3s gap. One failed-job retry at the same reviewed head is running (attempt2). Its dependency setup completed in13s, recovering the lost budget. No source/timeout changes and no local serving-node build.