Uplink: real Chrome control and whole12 restore pass; old timeout cause unknown
Runs: scratch/runs/20261112-halflife-chrome-endpoint-repair-1710/,
20261112-halflife-chrome-endpoint-repair-1710-control2/, and
20261112-halflife-chrome-endpoint-repair-1710-original/.
The frozen Chrome launch passed an actual blank endpoint/CDP/capture/clean-close
control before the original was started in the same box. The original then
verified the complete twelve-file VFS, loaded the earned crowbar checkpoint on
actual SoftwareGL, and demonstrated fresh movement. The historical
1655 startup timeout was not reproduced;
its causal failure remains UNKNOWN. No production repair is justified or
claimed. Missing DevToolsActivePort was the caller's final observation, not
an established cause.
Frozen launch diagnosis
All589 predecessor frozen files matched locally before allocation. The browser
source SHA256 is 18b0d0f6b957a71b494fb497023c479dcf41eddf1c4ca95f3629119717b866d5,
byte-identical to both1655 and the successful1635 caller. The custody collector
is unchanged f8e6e8be352a20f668185c907fdcf4308f5cb93f8c6b8297fada063fc8140b77.
The original emitted vector is /usr/bin/google-chrome with headless=new,
1024×855, no-sandbox, no-first-run, no-default-browser-check,
disable-dev-shm-usage, enable-unsafe-swiftshader, remote-debugging-port=0,
a unique user-data-dir and about:blank. There is no repeated unchanged original
launch: the blank control is first, then one original attempt.
The diagnostic launcher wraps the same collector only to measure its real Journal.write/fsync calls and record bounded process observations. It preserves ptrace returns, stop handling, durability, admission and cleanup. The control extracts the executable argument literal directly from the frozen browser, runs through the same Node owner-exec/owning-parent contract, and retains actual executable/cmdline/status/environment hashes. It never starts a guest.
Actual control Chrome PID7162 resolves through the /usr/bin/google-chrome
Bash wrapper to /opt/google/chrome/chrome, executable SHA256
22e4d5ef0c637a87b59a9baad42625aba3284dc03d4a5908748e253556318c97.
The wrapper adds flags; the full resulting vector and parent/TGID hierarchy
are in startup-analysis.json and recovered launch/process evidence, rather
than inferred from the requested command alone. Environment values are hashed,
not published. Actual original Chrome PID8549 has the same executable hash;
recovered/webgl/actual-host-launch.json.gz retains its vector, environment
hashes, wrapper hash and profile observation.
The control publishes a real endpoint, validates its HTTP version, obtains Browser.getVersion and SystemInfo.getProcessInfo, evaluates about:blank to42, captures a CRC-checked PNG, and receives Browser.close. It exits0/signalnull without TERM/KILL; all130 custody-owned tasks are reaped, outer wait0 with both EOFs. The control's combined startup/capture/close elapsed report is2402.283ms. Measured custody service totals are1505 fsync calls /0.746874s and1503 journal writes /0.826616s, with zero dropped samples; writes include fsync time. These measurements include scheduling and do not explain1655.
The1635 evidence reports Chrome151.0.7922.108, but retains no executable hash, full inherited Chrome environment or per-write service timing. The1655 inter-record gaps include storage, scheduling and other tracer work, which cannot be separated retrospectively. Empty stdout/stderr and absent exit JSON also cannot identify its first causal failure. No historical binary/environment identity or storage-cause claim is made from the successful control.
Whole12 and bounded onward inspection
The authoritative 1635 export remains
649713 compressed bytes, SHA256
f523dfeaf57ef0f4d4c470965d4f036b3e67affde4498f203f55650db204d809:
12 exact paths /2006166 decoded bytes. Strict base64, every size/SHA, ordinary
required mounts and all fresh VFS paths match before native Load. The same65
original media files and185 HTTP pins pass. Native menus select
OpenGL/Default/640×480, then the bottom earned UPLINK01:48 slot. Passive
postLoadReady is unchanged; later full captures wait through Loading.
d15-readyview restores the bay/crowbar at health100/armour0. d16-bound
anchors the living owner; d17-fresh-right visibly strafes right. Actual guest
and renderer module SHA256 both equal
04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17, and the
owning context independently reports software:true.
The small crate is already destroyed and is not attacked again. Forward,
right-side and short left-alignment probes inspect the tall structure and the
narrow metal-jamb lane (d18–d22). One native Space jump (d23) does not
establish passage. The original default.cfg binds Ctrl to +duck, but the frozen
key helper accepts neither Control nor chorded movement. Crouch was therefore
not dispatched; no live helper mutation or guest binding change was used.
Backing clear and turning reveal the other red-lit bay direction. After a
jamb-obstructed approach, d33-clear-to-right exposes the Maintenance Area door
and vending window. d34–d38 approach and test use; E and a distant default
Tab use do not visibly open it. d39-close-use uses Tab at the newly closer
position and visibly opens the door, exposing the familiar tool chest,
table and vending-machine room. This appears to return to the earlier supply
room, not establish an onward corridor or encounter. No enemy, health change,
new obstruction destruction or new earned successor is claimed. The close
read-only export matches every unchanged whole12 save file.
Evidence, staging correction and lifecycle
Original postflight passes474 recovered references,41 full captures,24
postbinding captures with all22 anchored fields,35 independently released and
acknowledged pulses, native Chrome exit0/signalnull, actual outer wait0/both
EOFs and369/369 custody tasks reaped. No cancellation or remaining/unreaped
task is present. All original capture hashes/ages, passive readiness,
independent releases and terminal no-replay rules are preserved. Unchanged
focused/build evidence is reused by hash in reused-focused-checks.json;
there is no new production edit or build claim.
An earlier staging attempt in the first run stopped on a local observer pin mismatch: the worker edited the observer after its upload pins were frozen. That was an avoidable worker error. No Chrome was dispatched, caller admission closed, the failure was retained and the box normally stopped. The corrected source was frozen separately before the second box was allocated. During control retrieval, a local verifier incorrectly compared decompressed bytes with the hash of an already-compressed original. The retained inventory allowed read-only recovery to continue with exact compressed hashes and gzip CRCs; there was no source/Chrome/input replay or actual artifact mismatch.
Both boxes had independent observable two-hour watchdogs armed and observed before allocation. Unique evidence was retrieved and SHA-verified before each normal delegated stop:
| Box | Final state / expiry | Normal stop receipt |
|---|---|---|
bx_h262x9ah |
archived / null; stopped17:07:31.069Z | stop_80e5af8b2fa22bfe, completed, forcedfalse/errornull |
bx_j3htstm9 |
archived / null; stopped17:21:46.881Z | stop_484005c3ae815d93, completed, forcedfalse/errornull |
Initial expiries were19:05:36.126Z and19:07:44.434Z. No box was deleted or resumed. The first local starter/manager/watchdog actually joined0 with EOFs. The second starter's actual tool wait returned143; its manager-parent wait record is absent and the cause is UNKNOWN. Its watchdog actual wait0 with both EOFs and confirmed terminal record is retained. Neither the successful remote outer wait nor the stop receipt substitutes for that missing local join.
Git is the source archive: the original freezes Git
eec0398723f94568c7d524e9a94e791ad407651f plus its relevant binary diff and
helper patches. Each new run retains final helper patches; the predecessor
pins reference its existing Git/patch recovery chain. No source/media archive,
guest console/state/config/AI/health/clock/position write, forced export,
worker launch or deployment was used. Sampled /home free space stayed above
1.5GB (lowest direct observation1,524,191,232 bytes); this is not a continuous
free-space claim. No RAM, parity, FPS, audio or full27 claim.
Next: retain whole12 as the checkpoint. Before another allocation, add and test ordinary Ctrl plus independently released jump/movement chord support in a new frozen caller, then inspect an actual unvisited route with fresh frames. Exact 1655 attribution needs a reproduced causal observation; neither ENOENT nor the new successful launch supplies it.