Caesar III compact audit output and actual capacity refusal, 2026-11-12
No box was allocated and no native game was launched. Assignment 2 granary/wheat remains unobserved. This task measured the closed 2320 parent, tested a smaller lossless audit representation against its actual files and localhost provision/caller fixtures, then stopped at the actual /home floor. The historical 1135 accepted transition is not new food success.
Evidence: scratch/runs/20261112-caesar3-compact-provision-food-2340/.
The parent writes every provision event twice: control.record appends api-events.jsonl; the provision wrapper appends the identical event to uploads.jsonl. Both closed files are already gzip, each 2,146 bytes recovering the same 16,502 bytes with raw SHA-256 23b39f2585982f59f76db541573d0402ef084b6c5633cb8157fbe18773f61f36. finalize-remote.py consumes the caller journal to identify launched owners. api_command_custody.Transport.provision calls the unchanged q3-provider-upload.py chunk assembler; its large-file path stages one local 4 MiB raw range before PUT, then removes it. That staging remains a distinct live provisioning liability; this task did not change the shared transport or claim to eliminate it.
All 67 closed parent files were inventoried before and after, including device/inode, allocated bytes, raw size and SHA. All remained unchanged. Of these, 42 logical generated JSON/JSONL metadata paths were encoded into one compressed metadata stream with 41 unique raw records, an ordered offset/size/SHA index, and an encoded/raw seal. No source, helper source, patch, binary, game media or screenshot content was bundled. Every logical metadata path was independently reconstructed byte for byte. The stream includes immutable manifest/media identities, original request/ACK/failure evidence and actual parent/EOF receipts; compression does not reinterpret them.
| Measured representation | Stored bytes | Allocated bytes |
|---|---|---|
| Original selected parent output, including existing gzip | 451,893 | 593,920 |
| New compressed stream, index and seal | 142,967 | 151,552 |
| Reduction | 308,926 (68.4%) | 442,368 (74.5%) |
These are actual emitted files, not an extrapolation to complete provisioning. The old artifacts were preserved, so no existing disk space was reclaimed: the retained alternate representation uses additional storage in this task's own directory. The measurement demonstrates smaller metadata output, not a deployment of compressed metadata into every lifecycle writer.
The own-prefix compact_audit.py also streams all caller events into a single gzip journal with logical provision/caller aliases. control.py retains the parent's request-pending and sticky-terminal callback behavior, replacing its event writer and removing its duplicate provision writer. A stream is accepted only after close/seal, encoded SHA/size, gzip CRC/EOF, raw SHA/size and complete JSONL checks. Failed or unsealed streams cannot become a provisioning pass. The full lifecycle in this candidate helper was not exercised or authorized for allocation after the floor failed; a future worker must integrate metadata consumers and freeze the final helper closure before any live use.
Focused tests invoked the actual production provision transport and emitted caller callback against a loopback-only HTTP peer. The success fixture performed a real PUT of a tiny synthetic input, executed the transport's actual generated final-file SHA reread in a local fixture directory, and retained all command-status and completion events plus a 600,001-byte output. Additional emitted fixtures covered HTTP 503 before upload, lost POST ACK, and a command completion explicitly marked stdout-truncated. Each failure remained terminal; a later command was rejected without another POST. Failed-request and lost-ACK fixtures ran zero commands and zero PUTs. No remote service, box or game was contacted. Truncated gzip, changed raw SHA, absent seal and partial JSONL were separately rejected. The test parent actually waited status 0 with separate stdout/stderr EOF. The final suite passed four integration cases and four recovery negatives.
All 694 frozen parent inputs and all 114 original media files were reread and still matched. Accepted deadline-safe teardown/input fixtures were reused by exact SHA from reused-accepted-fixture.json; no unrelated matrix was rerun. Production tools were unchanged. Tests apply to run-local storage helpers, so no production code/build change is claimed. The recorded Git source base is 88c64a6ef2d6ced473255e6344698a2e8ea8f2d1; the relevant production binary diff is empty and owned-inputs.patch preserves the three new helper sources.
After the final focused tests, actual /home availability was 1,490,595,840 bytes, 9,404,160 bytes below the strict >1,500,000,000 requirement. This existing capacity deficit exceeds even the measured metadata allocation reduction, and the unchanged 4 MiB range staging would add another transient demand. There was no allocation attempt, capacity waiting loop, RAM gate, browser, native input, fabricated continuation, save/config/resource mutation, or new food claim.
No new lifecycle needs shutdown. Parent bx_r7jt4s8b was neither resumed nor changed: its retained evidence states archived, expiry null, stop stop_afe107279120ccc8 completed unforced/error null. Its canceled rome1.wav upload remains UNKNOWN and unreplayed; one completed C3.555 assembly does not establish complete provisioning. Next requires actual capacity, then final consumer integration and a freshly frozen, independently watched private box before ordinary compact tutorial/Victory/Clerk/Assignment 2 food play.