Stronghold render-sized observer, 2026-11-04
Input: d371d379f11db7e4bba7c702d516976dae724800 and
the prior drain investigation.
Evidence: scratch/runs/20261104-stronghold-render-sized-observer/.
The helper capacity and first-fault repairs pass offline. The one authorized original entry failed during bootstrap, before Chrome, because an ownership receipt was written before its output directory existed. That ordering defect is repaired and reproduced offline. No second original entry was made. No fresh game scene, CustomGame callback, setup, preset, Start or world was reached. The historical input-drain cause remains unknown; no production engine defect is established.
Fixed observation budget
The observer now chooses 50,000,000 native boundaries before execution, within the existing 360,000 ms observation deadline and 720,000 ms browser deadline. The authorized upper validation ceiling is also 50,000,000. Nothing resets or extends either deadline or the boundary count.
The original initialized 800×600 RGB565 render requires 1,441,804 native entries. Retained actual successful 64-step host ACKs advanced the observer by 64,000 entries while rendering. A 1M budget therefore cannot cover even one valid complete render; a host ACK is neither a completed render nor guest message dispatch. The fixed 50M capacity holds 34 measured full render passes (49,021,336 entries), with 978,664 entries for message/return overhead. This is conservative capacity sizing for a render/message sequence, not a claim that the missing historical sequence contained 34 renders or those dimensions. The runtime, route and command limits can stop earlier.
Each native block increments a counter. Only baseline, callback/branch, scene/setup saved-return history and host annotations become records. The 2,048-record, 1 MiB byte, eight-frame nesting, exact owner/instance/memory, thread, immutable-code, input binding and consecutive ACK guards remain. Callback and return records are never sampled out or replaced by a suffix.
Executed capability and first-fault checks
render-composed.cjs reads the existing unchanged native render fixture by
path/hash. It preserves its original native function and initialized caller,
surface and lookup data. Its observed arm runs the actual serialized owning
Worker installer, original pre-instantiation observer, exact owning-target
attachment, pre-advance check and successful ACK annotator. The offline CDP
adapter transports real attachment operations; it supplies no native callback
or scene evidence. The fixture has no game input queue.
For both retained normal and compatibility modules:
- The explicit old 1M budget throws at entry 1,000,001, with only 1,000,000 entries forwarded to the prior import. The enclosing native run aborts before returning from the render. No continuation runs after that first fault.
- The default new budget returns from the unmodified native render. Padding to complete 64,000-entry ACK quanta gives 1,472,000 total entries, including the caller and stop loop; the render itself still has 1,441,804 entries.
- All 512 MiB page hashes, general registers, PC, stack, lazy flags, native PC histogram, render output and every completed run-command result match the trace-armed control. No guest code/data writes are made by the observer.
- Boundary, record and expiry negatives stop the actual composed execution. Additional executable Worker tests exercise record failure at a real native callback and expiry before a native block executes.
Previously wrapped() recorded a fault, detached its observer, and returned to
the native run. It now rethrows after preserving the first fault and pending
frames. The driver also checks the original observer deadline before every
advance, including the first advance with zero prior block records. This closes
the interval in which an idle expiry could have disabled tracing before an
otherwise unobserved advance. Full-history callback/setup fixture tests retain
their explicitly synthetic dependency returns; they prove instrumentation
capability, not the unknown historical dispatch.
Actual successful ACK annotations now obtain a step-matched page snapshot and include queue, active input, last input, held mask and input telemetry. Their native rows include the last native PC and active saved-return frames. These annotations pass through the composed driver in the capability fixture. No successful ACK was reached in the fresh original entry, so there is no new live ACK or dispatch claim. An empty host queue still cannot prove guest dispatch.
Single authorized entry and offline repair
Full preflight verified 2,429 filesystem pins, 1,856 unchanged original media
pins, 30 retained helpers, both modules and 1,016 HTTP responses with content,
isolation and cache headers. Media and modules are referenced by existing
path/SHA only. The exact MDK 20:53:05 local-heavy release preceded acquisition
of the sole local runtime flock. Runtime/bootstrap/command/cleanup reserves
remain 720/60/40/180 seconds, with absolute cutoffs 23:50/23:53 UTC.
An earlier wrapper-only guard rejected a short-lived shell-FD flock owner
before preparation or an original entry. Its log is preserved separately. The
corrected wrapper uses the tested long-lived flock ... bash ... ancestor.
The sole authorized entry at 20:57:43Z then failed with ENOENT writing
browser/runtime-ownership.json.tmp before the original fresh-directory
creation. This was an introduced helper defect. Cleanup ran and the entry was
not retried. The repair keeps release/lock validation before side effects and
writes the receipt immediately after the original output-directory creation.
The regression executes the actual composed main through its first real
filesystem writes: old ordering fails ENOENT, repaired ordering creates the
directories and receipt, and a second entry still rejects. No server or Chrome
is invoked by that regression. The earlier admission test had mocked save,
which is why it missed the ordering defect.
Node 1142062, start ticks 187307189, and wrapper 1141984, start ticks
187306734, exited with status 1 and are absent. There was no Chrome, game
server, profile, scene or gameplay input. The preflight HTTP server is closed,
its port refuses connections, and the runtime flock was released. Box
ID/state/expiry are NA; the coordinator retains lifecycle ownership.
Validation and evidence recovery
Focused local-runtime, owning-callback, original-bootstrap and route-recognition
tests pass, including actual old-order/new-order filesystem execution. Both
module native render/control checks and first-fault negatives pass.
flock /tmp/wine-build.lock bash tools/build.sh passes.
Normal module SHA-256:
78a0b1d61b23428a1ae909d739440ae0de059bbbd84a69a2ad3e4c37471ad664.
Compatibility module SHA-256:
99b579cd17b57f94145df92ae6d28110f24fc284d13a1b2c0ebe3596bd10a77f.
Original EXE SHA-256:
9024c533095f2227b0105b7e8a9e4cf28ec4fb206b165629787b2c97cf99f13f.
All 53, 554 and 523 artifacts in the directly relevant predecessor seals remain unchanged. Source recovery uses exact Git SHA plus relevant binary diffs and the owned new-note patch, without source/media/module snapshots or tars. The launch candidate's old local-runtime diff is preserved separately from the final offline repair. A future original requires a separate assignment; this task stops after publishing the tested repair and the bootstrap failure.