Nox corrected status and attributed ordinary host route
Evidence: scratch/runs/20261110-nox-corrected-status-host-route-0423/.
The run reuses fixture 682f77655 and readiness repair bce2b316f unchanged.
It preserves ca2989f62 actual-wire pending/predecessor attribution,
e519e4b99 command-scoped passive replies and capture-only intervals, explicit
reviewed cursor anchoring, and befbafc5b synchronous buttons. Shared providers
and other titles are read-only.
Status prerequisite
Before allocation, the new run-specific wrapper reused
status-offline-fixed.py with its inspect.cleandoc correction. The new
tools/nox-status-source.test.py extracts the actual wrapper's source expression,
reproduces the original line-2 indentation error, compiles the exact corrected
emission, and executes it against finalized actual files from the previous run.
Only its remote run path is mapped to the retained local files; the emitted
source is unchanged. All 37 retained artifact references are independently
SHA-256/size checked first. It observes startup ready and the actual failed
parent wait with separate stdout/stderr EOF. This test starts no browser,
sends no protocol, writes no guest state, and sends no process signal.
The exact newly emitted source SHA-256 is
b53c6c2508b177c0e51dbd81e1ab0bd89b5b06825801e1449907b06592f6c158.
Offline admission
The actual emitted passive driver passes eighteen controls, including its Worker handler and installer, actual host initialization, canonical launch module and installed observation instance. Four actual emitted CDP controls join the terminal pending operation to its actual sent expression and preceding reply, preserving the 20-second timeout and zero protocol after first failure. Four production pointer controls pass.
An initial ordinary-peer invocation used the passive driver and failed in its offline fake-browser protocol. Its evidence is retained. The ordinary peer requires its ordinary generated driver; that prerequisite is corrected before allocation. It is not native evidence.
All 27 matching ordinary controls passed before allocation, along with the
locked build. The native production module was
4f2c69a1528e01a9ab5127201d5e8a222da2a35236377a1351f444f55fa5a7d9
(2,186,099 bytes); the observation module was
716d28c256b2d204e83a3e73a3bcd843e1afab9d62c957f868b762faacef19f5
(2,186,872 bytes).
Git is the source archive: the run records its exact base SHA, relevant binary diffs and the owned new test patch. No source tree archive or media copy is made.
Native route and exact first failure
The corrected status wrapper succeeds live: its emitted source matches the
preallocation SHA above and observes ready with the owning parent alive. Escape,
the explicit move-only (0,0) anchor, No, Multiplay, a separately bound
capture-only settling interval, and Network all complete. Each returned image
is individually reviewed before dependent input.
Host Game also completes. All 34 command-scoped CDP sends have matching replies; its usable image shows the blue placement flag and “Place your game on the map.” The previous Host Game timeout does not reproduce. Its earlier underlying cause remains unknown; this run does not retrospectively attribute it to a record limit.
Host Game moves the original cursor from the physical reference (115,62) to
(408,239). Before another button, an ordinary move-only input to the bottom-right
edge brings both original cursor and reference to (639,479). That image and
passive state are reviewed. This uses ordinary physical movement and original
clamping; the synchrony guard and explicit anchor rules are unchanged. A reviewed
map click opens selected Conjurer with Accept visible.
Conjurer Accept is dispatched with synchronized buttons. Its returned screenshot
shows default Jack customization, but the command receives no successful ACK:
at 2026-11-10T04:44:44.238Z, capture completion rejects the passive reply with
reason: record-limit. The interval starts at global record 11,815 and retains
only 185 rows before the cumulative 12,000-row buffer fills. Its lifetime event
count is 12,956. This is a concrete observation-buffer failure, not a guest
failure or a pending CDP timeout.
The last completed actual request is 373, Runtime.evaluate, operation
nox-consumption, Worker session 46BB8707C9B5EE82ED36AF032CA4A305, target
6EAB63294FF3B0125E3A81A06D216051. It returns at 04:44:44.215Z.
The exact globalThis.__noxConsumption.finish(...) expression reconstructed
from the actual capture identity is 1,850 bytes, SHA-256
66fd00c80c8f5839b9dd922e1a186446c59e56459a58f48cc8fbe4976e40016b.
The terminal record has zero pending requests. The sent/reply trace has no
rows after the first failure.
The independent owning watcher sends SIGTERM at 04:44:44.266988Z, about 29 ms
after first failure. The local request wrapper also cancels owning custody. A
subsequently invoked local fetch helper refuses on its missing successful ACK
before any remote fetch; its duplicate cancellation attempt also refuses locally
on the immutable existing attempt file. Those local harness errors are retained.
There is no later browser protocol, input, capture, retry, quit, extension or
second native execution. The incomplete command's existing screenshot is recovered
only with finalized evidence and reviewed after cancellation. Default Jack's
Accept, Bunker and the staff route are not attempted.
Supported offline causal repair
The Nox recorder now retains one command interval in its Worker buffer. A new valid begin releases the previous completed interval, whose complete reply has already been returned and whose rows have been sent to the existing console sink. The 12,000-row capacity now applies to each command, rather than cumulative session history. Global sequence numbers, total record count and interval start indices remain monotonic; live replies remain command-scoped. Snapshot scope is also explicitly command-scoped, so it never claims to hold discarded history.
The lifetime 2,000,000-event limit and original absolute observation deadline are not renewed. Within-command overflow remains terminal; so do sink, identity, incomplete-call and deadline failures. An active unfinished interval cannot be overwritten by begin. No guest writes, artificial native calls, clock/tier changes, extra browser requests, timeout extensions or shared-provider edits are introduced.
tools/nox-passive-record-storage.test.cjs executes the retained actual emitted
installer and the repaired installer in real Node Workers, using the exact same
retained native canonical/observation modules and actual host imports. Complete
compiled x86 CALL/entry/RET intervals reproduce the old failure after 3,000
commands / 12,000 rows. The repair completes 3,001 commands / 12,004 rows, retains
only the last four rows, and emits every row in sequence. Installed instance
identity and deadline are unchanged. This is offline causal evidence; the
repair has no native validation in this task.
An initial added regression failed on a cross-context empty-array assertion; that test comparison is corrected. Separate stale-build/module refusals arose while another worker edited owned shared WAT files. Those files remain read-only here; failed offline attempts are retained. The exact retained-module comparison above avoids mixing native and subsequently changed modules.
Final repair validation
The locked build passes after the owned repair and tests. Fresh generation from that checked current source passes all eighteen emitted-consumer controls, including the actual Worker handler, actual host initialization and installed instance. The installed recorder completes 12,002 command intervals, emits all 12,002 rows contiguously and retains only its final row; its absolute deadline is unchanged. Four freshly emitted CDP attribution controls also pass. Compiled CALL/entry/RET controls complete 3,001 four-row intervals with balanced frames; within-command overflow, unfinished begin, identity mismatch, lifetime event limit and the original deadline still fail terminally.
This final offline build includes the other worker's concurrent shared WAT
changes without editing or committing them. Its production module is
1b7319ff24f64d6c886803f23cb9ea88447967da4ba88fef0b177961007dc608
(2,186,131 bytes), compatibility module
8e6c902e2d7fb9fc380c45f1cfade06c7f20a8a5a3267929dff13e3a595786fd
(2,188,777 bytes), and observation module
6311de81bea2171d4a84f9a92f7ad159fc13905eebbccf0530752a6fbd6228a8
(2,186,904 bytes). These are offline-only identities, distinct from the
native modules above. The run preserves separate Git base/diff records for
native execution and final offline repair validation.
Lifecycle
Only bx_75kn78v3 was allocated, UUID
7ed8dd32-6068-4512-960a-a597040bf456. The independent watchdog was prearmed
with immutable deadline 2026-11-10T05:04:22.035567Z; initial API expiry was
2026-11-10T05:04:23.671Z. Prelaunch verifies 866 inputs / 110,408,873 bytes.
Custody reaps all 337 owned descendants with complete kernel fork lineage, no unresolved lineage, and absent owned processes/ports. The actual owning parent returns 1 with separate stdout/stderr EOF. Chrome's separate exit JSON is not finalized; no separate Chrome wait result is claimed.
Before normal stop authorization, independent rereads verify 174 artifact references, 152 unique contents / 105,627,320 unique bytes, four finalized gzip streams and eleven full PNGs (chunk CRCs, complete zlib inflation, dimensions and scanline bounds). All eleven images are individually reviewed; the final incomplete command image is reviewed only after cancellation.
The sole delegated unforced stop stop_261edf34ed33fc8c completes at
2026-11-10T04:45:37.576Z. Independent API and receipt reads confirm archived,
expiry null, completed and forced false. The watchdog supervisor records actual
wait 0 with separate stdout/stderr EOF; watchdog and supervisor are absent.
No other box is operated on.
The supported repair is bounded command storage. Ordinary default Jack Accept, Bunker and staff progress still require separately authorized original execution; the earlier Host Game timeout's cause remains unproved. This worker releases its paths and stops after publication.