Pawn: bounded store observer controls; original cause still unknown

The run-local observer is implemented in tools/pawn-store-observer.cjs and exercised by node tools/pawn-store-observer.test.cjs. It repairs observation granularity and receipt interpretation without changing production primitives. Evidence is in ignored scratch/runs/20261113-pawn-store-observer-1630/. The fourth-history writer and selected PV-byte provenance on the original remain UNKNOWN. No original was launched or input sent in this task.

Read the 1550, 1520 and 1445 notes. The authentic 100-byte save remains at scratch/runs/20261113-pawn-modal-guard-1445/pawn-1445-secondmove.gam, with SHA-256 811659988ee98f19ca08615f1191431993f5d51646ea41efed5fdb6f0ff5cf77. Its existing e4/Nf6/d4 witness and fourth record retain their prior limits. No save, original guest, position, AI, configuration or clock was edited.

Observer contract and actual primitives

An exclusive controller supplies the exact owning exports and memory, one or two fixed guest ranges (at most 4096 bytes each), a live exclusivity predicate, and finite time, instruction, journal and epoch bounds. It must already have paused all other executors and host writers. A true predicate in the synthetic fixture establishes only fixture exclusivity. It is not an ownership proof for a running original. The helper must never be invoked through a concurrent control evaluation while the ordinary scheduler is executing that instance.

The controller arms native_store_arm, which itself clears decoded caches and enables instruction granularity, then uses set_rep_store_step to obtain the existing protected entry policy and one-element REP continuations. It records baseline bytes, complete receipt, thread, entry PC and publication serial. Each step() grants run(1) and samples the receipt immediately afterwards. Each range has independent epochs. Exactly one store produces a journal record with epoch delta, actual PC, address, width, before/after bytes and a same-value indicator. Two writes by one instruction without an intervening sample are UNKNOWN, even if the final value looks plausible.

The actual build wraps exports through tools/protected-writer-entries.js. native_store_get is deliberately denied while protected mode is active: its conditional returns are outside the structural read-only export policy. This is intended conservative behavior, not a primitive regression. The helper uses the existing native_store_share(0) seqlock publication and atomic reads from the owning memory instead. It verifies memory/receipt agreement before arming and refuses unstable publication. No production policy was relaxed.

set_rep_store_step also installs a value watchpoint. Watching the target data can introduce a debug halt after its first change. The block count includes an attempted entry and cannot alone establish committed execution. The helper watches entry code instead and rejects debug halts explicitly. It never patches that code or adds a guest instrumentation store.

Continuous scope checks include publication serial, receipt generation and ranges, bytes, expected EIP, thread and API count between owned steps. Loss of ownership, unsupported/trapped execution, invalidation, host stores, overflow, multiple writes or any finite bound returns an explicit unknown:* result. Unknown stops retain the observed prefix with complete:false. firstWriters means first observed since this baseline, never first ever or first since an unobserved allocation. Missing events do not prove no earlier writes. Raw host memory writes, including same-value writes, are outside the contract and must be excluded by the actual exclusive owner. A selected PV address also needs independent pointer/allocation-lifetime checks in that predicate; this helper does not infer dynamic addresses or their lifetimes from fixed ranges.

Meaningful interpreter controls and original coverage frontier

Nineteen controls use the shipped interpreter and authored isolated x86 code, with an unchanged notepad PE as a loader fixture. Arming/stopping preserves the fixture's guest allocation bytes and CPU registers. Controls establish:

The original Pawn executable is independently hashed to 21f78408e306d77103484bc90accfd511a141131ed1e14990380ecea19de5a58. Its exact writer at 0x402a4c is 8f 44 1d 00, POP [ebp+ebx]. Copying only these four instruction bytes into the isolated fixture demonstrates:

  1. Protected mode rejects this opcode before changing ESP or accepting a store.
  2. Instruction-only mode executes it and records the correct PC/address/width.

The protected decoder also excludes the preceding ADD forms and MOVSX at 0x402a38; src/07-decoder.wat explicitly describes its admitted subset. Thus protected continuous original-wide coverage is not qualified, even if disk becomes sufficient. Instruction-only mode alone does not supply the protected native/host entry guarantees. Do not silently downgrade protection, skip an unsupported interval, rearm and call the next writer the first writer. No production regression was demonstrated, so no production fix is justified.

Validation and remaining work

The locked production build and nineteen focused controls pass. Their child processes have actual wait results and separate stdout/stderr EOF receipts; the board watcher is explicitly reaped. Source provenance records exact Git, relevant binary diffs, new helper patches and actual source/module hashes. Git remains the source archive; no source trees or media were bundled.

original-decision.json records the fresh heavylock, actual-child and disk observation. There was no floor-only launch, watchdog, guest input, third move, UNKNOWN replay or original native wait. Native/watchdog waits are N/A, not successful runtime receipts. No independent cleanup or box was performed: box ID none, state NOT_CREATED, expiry null, stop receipt N/A.

Next requires a qualified continuous observation mechanism across the actual Pawn instruction/host boundaries, sufficient sustained disk headroom and a fresh source/module freeze. Any heavy original must hold the heavy lock, exclude actual runtime children and prearm an observable independent deadline guard. Review native e4/Nf6, then d4, before any original first-writer claim. Observe history and selected-byte addresses from a proven baseline with stable pointer lifetimes; preserve explicit unknown wherever continuity is lost.