Original LMarbles Puzzle8 manual attempt, 2026-11-18

The requested clear → onward → ordinary Quit sequence was not achieved. The retained run is scratch/runs/20261118-lmarbles-current-manual-puzzle8-clear/. Native exit 0 after cancellation is not ordinary Quit or puzzle completion.

Observed progress

Complete screenshots were reviewed before individually chosen finite inputs. Game freshness stayed 5 seconds; confirmation 500 ms, independent UP 1750 ms, observation/completion 3000 ms, and the finite watchdog/Quit reserve remained unchanged. Rejected intents were retained as NOT_SENT, followed by a new capture and a distinct intent when retried. Recent recorded expired ages included 5.016–8.170 seconds. Smaller whole-desktop previews and removal of redundant path text did not resolve the delivery latency.

Runtime and closure

The published d7a471f0a controller/common helpers were recovered exactly from Git, using the retained qualified module SHA-256 0d383140f6d4207d18860e3cb22fc8fcc47a5d260b8d5f2443222396b245fdd5 and its 251-source pinned graph at runtime Git 1c56d4135c03ee060c25cabcea031810f4f5f8dc. There was no production-code change, repeat compilation, or broad qualification rerun. The prior controller19/native2/locked-build results remain prior evidence.

Seven originals have actual native waits of 0 with both EOFs. One early enclosing invocation ended SIGTERM/143, and a preparation failure exited 1; these remain failures, separately from the native waits. Twenty supervision receipts report no leaks. The final runtime and its manual helpers exited 0. All recorded closure checks and their limits are retained in closure-audit.json.

One preparation scope bound a changing black startup image: the native paint check refused it, while permanently consuming the bind reservation. A separately recorded fresh bind was refused as permanent; the scope was closed without any gameplay input. The next original used one fresh menu bind successfully. No request was sent after that scope's recorded UNKNOWN closure.

Repeated actual disk-floor breaches prevented continuous floor compliance; the final original closed at 1,498,554,368 bytes free. Verified, reversible compression of older closed, untracked logs and evidence metadata recovered 413,184,000 allocated bytes across five passes. Every removal followed disk round-trip SHA/size verification, a stable-source check, and a durable recovery recipe. Held titles and active outputs were excluded. FD visibility was incomplete: no observed selected descriptor is claimed, not complete proof of no open descriptors. Subsequent shared growth exhausted that headroom again.

Selected screenshots and all unique native records were retrieved and verified before removing the owned tmpfs directories. Recovery recipes, source binary diff, owned new-input patches, selected-image inventories, and actual waits are retained; Git remains the source archive. No old endpoint is reusable.

No agents, workers, Wine reference runs, Python, public deployment, or test boxes were used. Box ID null, state NOT_CREATED, expiry null, stop receipt null. Clear, onward, and ordinary Quit remain open.