Q3 owning checkpoint caller integration, 1535

Run: scratch/runs/20261109-q3-checkpoint-caller-integration-1535/. Offline only. Historical1475 cause and browser EOF remain UNKNOWN. No native qualification, browser readiness receipt, or native retry is claimed.

Production change

tools/q3-current-module.cjs now has the explicit version q3-checkpoint-b993419a5-v1. It authenticates the published b993419a54d732a67696851d8f2dc50b574b625b observer, separate-stream helper, and owning launch client against Git bytes and requires that publication in both HEAD and origin/main. Only this version accepts the new literal Q3 prelaunch SHA. CPU, modules, custody, HTTP providers, startup admission, and every other existing provider retain their prior pins. The original default migration literals remain intact: unversioned preparation refuses changed prelaunch bytes instead of silently refreshing a frozen pinset. No observer feature is selected by default.

The preparer generates q3-checkpoint-caller.py beside the run's input files. It accepts only the exact retained1475 phase caller SHA 5b52e6438e2ec3f9676d11584dd52c5a435b1b17412c4ac2d5c1dec33301bb1e. It changes that actual caller's binder to the published observer's bind_remote, and its cancellation checkpoint branch to observe_checkpoint(task, session, session.remote). It waits for the existing owning launch to start before admitting that observation. The original physical proof gate, cancellation request, original-command branch, and postlaunch evidence checks remain present. Historical callers are unchanged.

The generated caller, its explicit source input, and the run's control.py, physical_gate.py, and cancellation_gate.py must all be in the pinset. The observer's transitive Python dependencies are also included. Prelaunch checks the explicit version, literal provider hashes and lengths, generated caller and run-local dependency membership, compiles the caller, and actually imports all three Python providers, including the stream helper's deferred client. All this happens before the runtime spawn. Default-off prelaunch retains its existing execution path and separate stream drain.

Example preparation, with small run-specific inputs already in an owned run directory (this is preparation only, never allocation or launch):

node tools/q3-current-module.cjs --prepare scratch/runs/OWNED/prepared \
  --checkpoint-observer-version q3-checkpoint-b993419a5-v1 \
  --caller scratch/runs/OWNED/phase-input.py \
  --input scratch/runs/OWNED/phase-input.py \
  --input scratch/runs/OWNED/control.py \
  --input scratch/runs/OWNED/physical_gate.py \
  --input scratch/runs/OWNED/cancellation_gate.py

Additional runtime/run inputs still require explicit --input entries. This does not authorize a native attempt or reuse historical success evidence.

Verification

production-cap-final.log records the production CLI preparing the pinset and caller, an executable import of the entire generated caller, executable prelaunch checks, and the generated phase's actual launch() function with the real owning Session, actual generated parent and actual prelaunch. Only the outer transport and runtime are inert. No Boat, browser, game, or native custody supervisor is run.

Six actual caller cases pass: partial checkpoint, damaged checkpoint, prior failure with a valid checkpoint, cancellation with a valid checkpoint, diagnostic journal cap, and a valid checkpoint reaching the unchanged physical gate. The offline physical gate deliberately refuses native proof. There are no guest commands or successful native cancellation receipts. All created local transport children are actually waited; stdout and stderr EOF remain separate. Started runtime cases have nested actual waits; the cap case uses the helper's explicit diagnostic-failed-after-actual-wait journal record, not a fabricated normal receipt. The normal prelaunch-child receipt is absent in that case. The final subreaper sweep has zero orphan children. The independent 120-second sibling watchdog was observable before tested children, did not fire, and was actually waited separately.

Nine executable prelaunch cases cover success/default-off and refusal of a missing/stale observer, missing stream helper/client, missing generated caller, missing caller-local dependency, and unknown version. Production preparation also refuses six stale/missing provider read substitutions, a missing caller dependency, and changed caller source before producing outputs.

observer-final.log passes the retained1525 actual generated-parent controls: spawn failure, missing/partial/damaged checkpoint publications, prior failure and cancellation wins, absolute deadline expiry, journal cap cancellation, FIFO/symlink/size/gzip refusals, and real separate stream waits. Its retained1475 exists-only poll still demonstrates old false-ready behavior for partial JSON and a prior failure. This does not identify the historical native cause. The observer retains the same 60-second observation and 70-second transport cap. Browser startup deadlines and cleanup budgets are unchanged.

Focused existing controls pass: native launch client; actual stream EOF and inherited-writer cases; prelaunch import/barrier controls; Q3 startup; original current-module controls against frozen prelaunch bytes. The old stream test initially refused its missing current-module input; stream-fixture.patch records the sole in-memory addition required by its fixture, with all original assertions retained. Existing test files were not edited. Earlier failed test iterations are preserved, including a transport-wrapper cleanup failure and an overstrict normal-receipt assertion for diagnostic-cap failure. Neither is promoted to success. The final transport makes Session directly own the generated parent; it does not extend a timeout.

The final locked build is build-final.log. Evidence records the exact Git base plus the small owned patch; no source tree, media, or module copies were made. Large owned JSON evidence was losslessly gzip-compressed with verified SHA/length mappings in lossless-json-storage.json.

Ownership and lifecycle

Existing providers remained frozen until Q21495's final published RELEASE STOP at 14:18:04Z, commit f017d680e24167b7e95e824ddbc993278b404fd6, after its exact UUID fcb86e70-6cdf-4166-8a5c-c03951aa45fc archive and physical drain. Fresh OWN for these two Q3 providers followed at 14:18:46Z. Q2, custody, CPU, WAT, host, Redline1515, observer, launch client, stream helper, and startup files were not edited. No new worker, deployment, Heroes II or CLAUDE change occurred.

New box ID/state/expiry/stop receipt: NOT_CREATED / N/A / N/A / N/A. Inherited boxes were not contacted. A future native attempt needs separate authorization and fresh versioned preparation; historical browser EOF and historical1475 cause remain UNKNOWN.