Diablo II XP401: whole-stream telemetry measured; capacity still prevents allocation

The new run-local telemetry callback stores the actual 0140 parent's 734,390 retained bytes as 93,555 bytes including its verification seal, saving 640,835 bytes (87.3%). All 1,248 events and every nested body are preserved byte for byte. The comparison inventories actual on-disk encoding: that parent has one raw telemetry file and no compressed prefix. Existing compressed prefixes would count at their encoded size, never their inflated raw size. This read-only historical telemetry benchmark sends no API commands and rewrites no parent files. The saving reduces the new benchmark's retention; it does not reclaim 640,835 bytes from the protected historical run.

Evidence is scratch/runs/20261113-diablo2-compact-0315/. telemetry.py holds one continuing gzip stream per caller process, with flush/fsync after each callback. It retains the original JSON serialization, ordering, field values, terminal side receipt and all inherited Transport behavior. Process-specific streams prevent manager/watchdog interleaving. Normal caller exit closes and seals each stream. Seals bind the encoded and decoded SHA-256/size, schema, record count and path. The actual caller's read_telemetry() validates the seal, gzip footer, complete JSONL and decoded identity before returning events. An interrupted, missing-seal, modified or truncated stream remains unqualified; the reader refuses it without repair or replay. There is no live telemetry reader in the inherited provisioning, input or recovery callers that requires an unfinished stream. Their terminal admission still uses the original side receipt and original Transport failure state.

Ten actual local HTTP loopback cases exercised both production Transport and the inherited exact Envelope through the new control.record callback and caller reader: success, failed submission, output truncation, UNKNOWN outcome and disconnected status for each. Each failure retained terminal admission and refused a second command before POST. Raw/gzip event consumers were exactly equivalent; missing seals and damaged gzip were refused. Two independent real subprocess callers also emitted 40 events, sealed through the actual atexit callback and joined with code0 and both EOFs. Their original fields, bodies and per-caller order matched exactly. These are local transport and storage tests, not native game input or gameplay evidence. All inherited control functions except the callback remain AST-identical; no ACK, deadline, release or uncertainty rule changed. The per-record gzip approach from Darkstone27dc1a8fd was not reused.

The capacity blocker remains larger than this telemetry saving. Completed resource0250 exhausted all six authorized eligible raw rollouts, reclaiming 4,071,424 allocated bytes before its evidence overhead, but measured only 1,480,351,744 available bytes at its verification. The current run's fresh exact availability and refusal are in capacity-decision.json; it is still below the strictly-greater-than 1,500,000,000-byte requirement. Separate resource0335 and dense0340 remain coordinator-owned work; this worker performs no independent cleanup and credits no prospective capacity. No box, provider upload, guest write, movement, quest input, exit, combat or successor save was attempted. This is the measured limit of the authorized telemetry improvement, rather than another allocation that loses the floor during gameplay.

The published startup reader875726057 was reused without its old matrices: 21 sealed artifacts and 65 original save/media/module/helper/proof references passed independent rereads, along with provider5c2's published bytes. The actual startup reader exited zero with both EOFs and the exact whole-four XP401 save dependencies. The original 187,872-byte renderer was recovered and verified in RAM from Git blob 0ad94648f373fa133815cf0e83958ab4a9ce6743, reachable at a04a489fc677073dbb569f6aad4efbfbca558318, SHA-256 4235d20d7fc2404b38f6e4f7f80c92d5f94f1513b4d6d7dcf04be34162a4b087. It was not written into the shared tree or repinned. Parent0140's entire sealed evidence inventory also passed a final reread unchanged.

Only owned ignored helpers and this note changed. Exact Git SHA, relevant binary diff, independently reconstructed helper addition patches, identities, tests and a reread evidence inventory are retained. No source/media archive, history/DB/cache/reserve mutation, worker, public deployment or production change occurred; no production build is claimed. The owned board watcher was stopped or confirmed absent, with original-parent wait/EOF uncertainty retained.

Box ID none, state NOT_CREATED, expiry null, stop receipt N/A. Historical bx_hca9hq83 remains untouched; parent0140 reports archived with completed normal unforced stop stop_79836825abd83106. Its original title, normal Load/camp and XP401 stats remain inherited evidence only. Fresh native movement, camp exit, combat, retreat and earned successor save remain unproved.

Next requires additional actual authorized capacity. Before any new allocation, prepare the valid full descriptor and freeze final helpers, observe an independent immutable watchdog, and recheck the exact original523/current04/ provider5c2 closure with explicit RAM ranges. Then use individually reviewed fresh full frames and independent UP for native quest, movement, camp exit and short safe combat/retreat. Only native earned progress justifies SaveExit/title/all-four export/fresh normal Load and movement. Uncertain old inputs and the refused parent quest are never replayed.