GeneRally bounded controller independent rejection
Historical source last updated, 2026-10-27. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
The candidate change adds a serial loading controller with bounded chunks, exact completion ACKs, frozen-zero snapshots, a durable pending/failed cursor, one global deadline, and post-loading capture/input gates. Those mechanisms do not establish exclusive ownership of the actual mailbox: the CLI locks args.state + '.lock', while Mailbox.__init__ only creates the requests directory. exchange checks collisions for the selected ID, not another owner's outstanding request.
The independent actual-CLI test starts an owner at next ID12 and withholds its snapshot reply. The owner stays alive with a durable pending cursor. A competitor using the same cursor exits1 without dispatch, demonstrating that the existing lock is effective for that path. A competitor using the same mailbox, an alternate cursor path, and next ID13 creates requests/0013.json while request12 remains unanswered and its owner remains alive. Both cursors are then pending. Zero ACKs were created, and no positive completion was fabricated. The test exits1 on this violated mailbox ownership contract. All three child processes are reaped.
This is a demonstrated source/CLI ownership defect. It proves that a state-path lock does not serialize the actual mailbox and cannot prevent later requests from an alternate cursor after uncertainty. The smallest repair scope is transport/mailbox ownership and its durable handoff policy across alternate state paths; this assignment does not implement that repair. No guest consumption or input delivery follows from a queued ACK.
Behavioral verification began with the unchanged historical {"action":"step","n":100000,"id":12} reproducer. The fixture validates the full pinned test/run.js SHA256 6dc104b3c10db701c0b1aec5954dd00cc29a7b6009970a5876ccc27426e7fa9e at authority 1e4fdb06f64e5f14ddec582111dfe7ecb71df6ba, blob f52b03a096a9a614b681cdf0f912ce20ca19fcba. It executes the unmodified contiguous production controller/stdin closure at bytes389135..405094, SHA256 96de2e3dbd35b5059e05e7a346a65f2926d643acf03b7af5cd9760255d1e2bf7. Host observations and batch cost are stubbed; actual step credit/completion accounting and ACK serialization are retained. The same request has no ACK within the retained synthetic0.12s allowance. This is protocol reproduction, not a guest timing measurement or new initialization/gameplay acceptance.
Startup and the final reproducer driver both use installed /home/user/.nvm/versions/node/v24.18.1/bin/node, --max-old-space-size=512, the same controlled environment and inherited limits. Fixture-recorded executable, flag, limits and boundary hash match the startup/source receipts. The sole reproducer-driver edit resolves the installed Node pathname; the reproducer and fixture remain semantically unchanged. No AS or DATA cap is applied to Node. CPU90s, wall120s, file24MiB, output2MiB, aggregate RSS768MiB for Node or128MiB for Python remain enforced by the trusted supervisor and shared single-permit pool.
The coordinator-verified predecessor manifest b329730100c9963b883269322fab05d75d2c0704741ac4b74d4e6e4533dbe81b and its709 payload integrity results are reused. Retained exact-current-source ACK fault, pending/poisoned cursor, deadline, nonmultiple-total, terminal/EOF and post-uncertainty input groups are historical qualified results, not rerun or promoted to independent acceptance. Partial/stale ACK and consumed-reply mutation challenges remain independently unverified: the new ownership rejection stops broader acceptance. The original failed nominal100000 synthetic0.5s case and corrected positive-only38s response/20s global result remain preserved and distinguished in the copied reconciliation. No generic receipt audit or unchanged passing group was repeated.
python3 -B test/independent-generally-bounded-acceptance.py --evidence /OWNED/fresh-evidence
Exit1 with the ownership diagnostic is the expected result on this rejected candidate. A passing repair must retain the same-cursor control and prevent the alternate-cursor dispatch without resetting either cursor or weakening the oracle. Future repair acceptance must independently cover this actual mailbox path and durable uncertainty before a real runtime is justified.