Uplink: actual native Control/chords work; wooden clearance still blocks onward
Run: scratch/runs/20261112-halflife-native-crouch-clearance-1730/, with
-control/ and -control2/ lifecycle evidence.
The new run-local requester and actual emitted consumer support native Control and finite movement/jump chords. Fresh original whole12 Load, crowbar, actual SoftwareGL and movement pass. A full frame during Control alone shows a lower viewpoint than the released view. Crouched forward and lateral movement, a new crouch-forward-jump combination, and a central approach under the wooden frame do not establish an onward corridor or encounter.
The first specific route blocker is the remaining wooden back/end face: central crouched movement closes on it, while the sign-wall/metal-jamb side approach ends in a close wooden pocket. This is an observed limit of these alignments, not proof that the structure is indestructible or every route is impassable. Health100/armour0 persist. No useful new checkpoint was earned; no native Save or successor qualification is claimed. The close read-only export matches all twelve unchanged checkpoint files.
Actual emitted input and controls
input.cjs, transport.cjs, drive.py and browser.cjs are new run-local
revisions, not production repairs. The requester validates complete key/chord
requests before publication; the consumer independently validates before
admission or dispatch. Chords require Control first, at most one W/A/S/D
direction, optional Space, unique keys, and bounded50–2000ms holds. Control
alone uses ControlLeft/VK17. Ctrl chords omit text and emit modifier flags.
Every key has its own prearmed release timer and bounded acknowledgement;
UPs are scheduled in reverse order independently. A failed/UNKNOWN DOWN stops
later DOWNs; all scheduled UPs are attempted before the group reports terminal
failure. Cleanup UPs can bypass admission without creating a new held key.
Native-group CDP failures defer terminal cancellation until release attempts
settle. No later command is admitted after UNKNOWN.
consumer-test.cjs extracts the actual execute body from the emitted browser
and drives the actual requester/numbered consumer, rather than a copied model.
Fifteen cases pass: Control, Control+W, Control+Space, three-key movement/jump,
held capture, capture failure, partial DOWN rejection, UNKNOWN DOWN, failed UP,
persistence failure, invalid key/order/duplicate/hold and zero-duration chord.
Invalid chords produce zero native effects. Tests verify DOWN ordering,
independent per-key UP attempts and sticky terminal admission.
Actual Chrome repeats the emitted-body cases with native CDP events. All DOM
edges are trusted and match the emitted ordering; a real Worker receives the
same event sequence. This Worker is an event echo control, not substituted game
logic. The original game's separate bound guest/renderer Workers and visible
crouch/movement provide the game evidence. The first local fixture assertion
compared DOM keydown with CDP keyDown; that test assertion was corrected.
Its Chrome was terminated and later profile-process absence checked; its
individual Chrome join record is missing. The final local control has Chrome
exit0 and actual parent wait0 with both EOFs and a joined prearmed watchdog.
A new passive in-hold full capture keeps the ordinary full-frame screenshot path and before/after22-field identity check while admission owns the chain; it does not replace the reviewed/latest frame. All eight original held captures complete before their independently armed release deadlines. This establishes physical input/capture intervals, not a simulation-step or keyboard-poll count. Actual UP attempt timestamps differ from planned times by−1..+2ms; both−1ms observations are retained rather than described as exact-duration holds.
The readiness, postLoad readiness,22-field identity/capture/admission and save
collector revisions match1710 by SHA. Their existing focused negatives/build
receipts are reused by exact hash in reused-checks.json. The ACK transport
retains numbered, sticky no-replay semantics with new complete-plan validation.
All new JS/Python syntax, staged import, final616-file pin verification and
actual provisioned-data/import gates pass before the original starts. No
production code changed, and no new production build is claimed.
Whole12 and native clearance observations
The authoritative 1635 export remains
649713 bytes, SHA256
f523dfeaf57ef0f4d4c470965d4f036b3e67affde4498f203f55650db204d809.
Independent strict base64, exact path set, every size/SHA, ordinary required
mounts and fresh VFS checks cover 12 files /2006166 decoded bytes. No older
auxiliaries replace members. All589 predecessor pins match locally,616 final
runtime pins match remotely, and185 HTTP pins pass before game startup.
Native menus select OpenGL/Default/640×480 and the bottom earned01:48 slot.
d12 performs normal Load with unchanged passive readiness; d13/d14
remain loading frames. d15-readyview restores the cleared bay/crowbar.
d16-bound anchors the living owner and d17-fresh-right visibly strafes.
Actual guest and owning renderer modules both hash to
04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17;
the actual owning context independently reports software:true.
d18-control-alone-held and released d18-control-alone show the lowered and
returned viewpoint. d19 crouches forward toward the tall face; d20 crouches
right into a close wooden recess. d21 turns toward the sign-wall side.
d22 crouches along that narrow lane and ends at its wooden end face.
d23 adds the newly supported Control+W+Space combination at that changed
position; the result is still a near wooden surface, not an onward view.
d24/d25 back clear; d26 inspects the opposite edge while crouched.
d27 approaches centrally under the open-front wooden frame and again ends
at its back face. Early in-hold frames need not include the full movement
later visible in the released frame; they are not extra distance proofs.
The already-destroyed small crate is not attacked again. No E/use loop, unchanged Space-only jump, familiar-room reopening, console/config/state/AI/ health/position/clock write, forced export or replay supplies these results. No new room, enemy, destruction or pickup is attributed to the close recess. Next: retain whole12; resolve the remaining wooden back/end obstruction with a newly justified alignment or ordinary interaction, or establish a different route from fresh full frames. Another unchanged crouch/push loop is not evidence.
Refusal, evidence and lifecycle
The first box's blank endpoint/CDP/full-PNG/clean-close control passes,133/133 custody tasks reaped. Its final read-only pin/import command is rejected by the provider with HTTP400 before any original browser starts. The emitted attempt is177897 bytes; excessive command size is an inference, not a provider-proved cause. Admission closes and the box normally stops. An earlier local adapter invocation without its controller module path also failed before API dispatch; subsequent requesters explicitly set that path. The corrected command reads a SHA-pinned manifest in place, is bounded below4096 bytes, and passes locally before a fresh caller/box. The second actual blank control passes with 131/131 custody tasks reaped before the original starts in that same box. The historical1655 endpoint cause remains UNKNOWN; no runtime endpoint repair is claimed.
Original postflight verifies365 recovered references,36 full captures,20 postbinding captures anchored on all22 fields,23 balanced acknowledged pulses and73 trusted events. All20 keyboard DOWN/UP pairs balance, including8 Control pairs. Original Chrome exits0/signalnull; the actual outer parent waits0 with both EOFs and364/364 custody tasks reaped, no cancellation or remaining tasks. Unique evidence is SHA-verified before delegated normal stop.
Both boxes have observable independent two-hour watchdogs armed before allocation. Both local managers and watchdogs actually join0 with both EOFs.
| Box | Final state / expiry | Completed normal stop |
|---|---|---|
bx_a6ynygpq |
archived / null | stop_de374c9c29e15fd3,17:37:21.165Z, forcedfalse/errornull |
bx_99b5q9pf |
archived / null | stop_7fd3882dc824d0ea,17:49:34.738Z, forcedfalse/errornull |
Initial expiries are19:31:45.795Z and19:37:56.317Z. Neither box is deleted or resumed; neither is left billing idle.
Git is the source archive: final original freeze records
700fbb1ec61fbbea6f27fdcbed74b81bd2cd4522, relevant binary diff and helper
patches; the final helper patch also preserves evidence-only readers/verifiers.
Exact compression journals retain raw/gzip hashes and roundtrip checks;
byte-identical owned files are deduplicated with hashes and every path retained.
Retrieval manifests retain original paths/hashes; a later exact gzip may replace
a local raw JSON, so readers decompress its sibling .gz before checking the
original receipt. No source/media tree archive is created.
The native-command disk samples remain above1.5GB. After ordinary quit ACK at17:48:31, shared free space is observed at1499824128 bytes during retrieval, then1497624576 during cleanup. No further native input is sent. Own exact compression/dedup and another worker's documented cache cleanup subsequently restore sampled free space above the floor. Continuous floor compliance is not claimed. No RAM gate, workers, public deployment, HeroesII/Claude panes, parity, FPS, audio or full27 claim.