Half-Life: bounded media caller prepared; fresh attempt held locally

Run scratch/runs/20261112-halflife-bounded-media-2125/. This investigation identifies the upload-path mismatch in e27fb3409 and tests a replacement caller locally. A fresh remote transfer and living combat consequence remain unproved because actual /home free space stays below the owner's floor.

The actual 2055 provision.py calls provision_files, which sends each entire file in one PUT. Its provision-blocker.json, stderr and control upload events record the 79,150,544-byte valve/pak0.pak body-write timeout at 30.000812 seconds, after 13,631,488 bytes of complete chunks. The uploader actually exits 1 with both pipe EOFs. This is an UNKNOWN partial PUT, not a failed chunk-assembly transaction. It is never replayed here.

The supported provision_chunked_files path already exists in tools/q3-provider-upload.py, SHA256 29d148d81929464f92bce963f68118337cf89524f0d3b77bff629ef12c9afa95. Q3's retained 1910-small-upload and 2055 callers select HTTP_PART=1048576 on that function's globals. The default is 4 MiB; the 1 MiB selection changes neither the absolute 30-second PUT cap nor the assembler's validation. tools/api_command_custody.py, SHA256 c5595d0d7b1084faa40bd4e0dffed08d24add35617bdae6b0e97d497d983c8e8, provides one-shot command identity/completion observation. Its default Transport.provision uses the chunked helper with the default range size; the qualified Q3 caller explicitly selects the smaller size.

The new run-specific provision.py freezes both tool hashes, selects the same 1 MiB path, checks the caller's effect gate before each upload authorization and remote command, and makes a transfer failure terminal in the transport. It does not import or reuse the terminal 2055 lifecycle. It requires the fresh caller to prove its outer process wait, both pipe EOFs and staged import before original launch. No production defect is proved and no production source is changed; no locked-build claim is made for this documentation-only commit.

check-transfer.py tests this exact adapter against a local HTTP receiver and actual Python assembler children with 2.5 MiB of synthetic bytes. The three PUTs are exactly 1,048,576, 1,048,576 and 524,288 bytes, with unique paths and independently checked range SHA256 values. Whole assembly and final descriptor reread pass. Deliberately corrupting the second received range refuses assembly, publishes no destination and emits no all-remote-verified barrier. Both cases retain actual child waits and separate stdout/stderr EOFs in transfer-check.json. All synthetic inputs and receiver files are removed. This is local transport evidence, not remote delivery or gameplay.

Read-only verification passes all 954 inherited upload pins. The actual pak has SHA256 c9eac1391845d6fabd93d7a1cc48281275410d35e01b74bd7f02325c65c99a42; pak-range-plan.json records hashes for its 76 ordered 1 MiB-or-smaller ranges without copying media. All 13 earned1825 save records independently decode and hash, totaling 2,342,063 bytes; the compressed original remains 758,472 bytes, SHA256 df8d61b2fcdd285a6a28e2e180930e6ee60eff388cd9bf31c557d3409dd18e51. The corrected lib/renderer-input.js is byte-identical to ae83f8d55. Browser, finite input/cycle, semantic delivery and capture selector helpers remain pinned by SHA. Solved control matrices are not repeated.

Local disk samples remain below 1,500,000,000 bytes; the final preparation recheck is 1,483,886,592 bytes, and the closing hold is 1,481,863,168 bytes. No one-shot API request, allocation, watchdog, upload, remote command or native input occurs. Root catalog RAM and reserve remain untouched. No current CPU/owning GL/full22 or original PAUSED frame, attack, hit, defeat, guard consequence, death, native Save/export/load or useful successor is claimed. The original13 is preserved. Git SHA, relevant binary diff and owned helper patches are retained; no source or media archive is made.

No box is created, so there is no new expiry or stop receipt. Prior bx_qwekfmpb remains untouched: its retained terminal receipt says archived, expiry null, stopped 2026-11-12T21:02:48.634Z, stop_361519b9f33a7197, completed, forced false, error null. This is historical evidence, not a new lifecycle query; its partial PUT stays UNKNOWN.

Next: obtain actual disk headroom, freeze a fresh complete provision manifest, prearm and observe the independent two-hour watchdog, then use the tested bounded caller on a fresh private box. Require actual provisioning join, all chunk/whole hashes and staged import before exact whole13 restoration, ordinary UI 03:18 Load and full native PAUSED review. Continue only one visually aimed finite action and full consequence per ordinary resume/pause cycle, with independent UPs. Transfer success alone earns no gameplay claim.