MDK2 bounded ordinary startup provenance

The owning binding now retains passive evidence from actual ordinary native run(budget) calls before observer admission. This implements the measurement recommendation in the checkpoint diagnosis. A host step remains a scheduling opportunity, not a fixed native instruction prefix. This change does not qualify a new checkpoint or authorize another original attempt.

__mdk2BoundTrace.report().startup contains at most 128 recent native run records, 16 power-of-two run-ordinal milestones, and 128 existing browser Worker sliceDone messages. Eviction counts and cumulative captured run/completion/block counts remain available. Milestones share captured records; they are not reconstructed history. Each run retains the actual forwarded budget, whether the original call returned, completed blocks, halt, measured elapsed time, and before/after registers including yield and guest tick count. Its owner identifies the authenticated transferred Module hash, compilation epoch, target, Worker URL, endpoint and sole owning instantiation. The existing binding rejects replacement instantiations.

After each run, the journal reads 48 raw words (192 bytes) at the actual ESP. Thread1, main loader identity, integer registers, aligned main-stack bounds and every translated byte must validate. Invalid context is marked invalid; the first capture error is retained. These saved words are caller context, never a pending-call ledger, CALL/RET witness, or GC entry. A valid record ending at a maintenance seam can establish that sampled PC and stack were actually reached. It says nothing about unobserved instructions between slice boundaries. Dropped records cannot be recovered from milestones or totals.

Incoming ordinary slice sequence/budget metadata correlates native runs with the unchanged browser Worker's outgoing sliceDone body. The latter preserves its actual ms, blocks, PC/ESP, yield and sleep fields, including the elapsed interval consumed by the adaptive host. The hook forwards the identical message, transfer arguments, receiver and result. Completed request tags are cleared; overlapping requests and unmatched responses remain explicitly uncorrelated and preserve the first diagnostic error. It does not consume get_sleep_yielded, which reads and clears a flag. Node's shipped Worker uses parentPort instead of browser postMessage; its report explicitly marks browser feedback unavailable. Native run timing and registers are still captured there.

The elapsed run() interval excludes the following caller readback. The Worker's enclosing elapsed interval includes instrumentation overhead, as does the adaptive host's normal feedback. Consequently this instrumentation cannot promise an unchanged adaptive trajectory or replay of either retained original. Fixed-budget native fixture parity establishes unchanged guest effects for identical supplied commands; it does not establish wall-time or original gameplay equivalence.

The journal seals on first native run failure, observer admission or terminal cleanup. Later reports and cleanup cannot replace its first failure/error or restart collection. Reports return copies of retained journal data. Capture failures do not replace the original native return value or exception. Malformed transport metadata cannot append unbounded objects. Production admission, the seven caller guards, scheduling, clocks, input, readiness, and the observer's one-shot 50M-boundary/360-second/512-record contract remain unchanged. In particular, 4641c7 remains unsupported. No saved return seeds an observer frame.

Evidence is in scratch/runs/20261105-mdk2-startup-native-provenance/. Both Stronghold0250 final 666b5ee4c and ReturnFire0305 final 691020294 runtime AND file releases preceded shared edits and executable checks. release-gate.json independently rechecks six exact PID/start-tick identities, four refused ports, removed profiles, normal Chrome exits, absence of heavy processes, both free locks and disk space. Source-only proposals were prepared before that gate without applying shared changes or running modules/tests.

Validation passed:

binding-ledger.json names the old/new integration and focused-test hashes; the local helper, resolver and observer hashes are unchanged. Historical inputs remain untouched. admission-unchanged.json verifies the complete createTraceDriver source against pre-edit Git 6910202949bf56e9b3488587fdd0b96893e07922. The final public-entry proof binds integration SHA-256 62e3c2ff1654d4808170c6ee47e577fa6dd0636510b952bcdc341c5e891bba9f. Evidence uses Git SHA plus relevant binary diff and owned new-file patches; media/modules are referenced by existing paths and SHA-256, without copies. No source tree snapshots or archives were created.

The next actual-native evidence needed is an explicitly authorized ordinary startup with this exact binding: correlate genuine budgets and Worker elapsed intervals with register/caller witnesses at an actually reached candidate seam. Preserve unknown or evicted intervals. Any later admission proposal must still justify absence of relevant pending tracked calls and continuously witness genuine GC entry/return using one unchanged contract. No original, browser, gameplay attempt or box lifecycle action was made here. Box ID/state/expiry: NA; shutdown remains the coordinator's responsibility.