Jedi controller: explicitly reviewed neutral reference

The source-only prerequisite is repaired in retained helper scratch/runs/20261103-jedi-controller-neutral-reference/neutral_reference.py and adapter bounded_reviewed.py. No game campaign was run. A future campaign requires separate coordinator dispatch and integration using the retained RECOVERY.md; the historical neutral image cannot authorize live input.

The actual predecessor step.py reads previous-state.json and passes its location into bounded.run, ignoring the later neutral state returned by the controller. Replaying the byte-exact retained controller reproduces its real western-corner stop before a fifth keydown:

Memory boundary, UTC 2026-11-03 Batch X Fresh distance
Immediate post-release 13:39:30.286 457411 −1.798464060 0.121526241
Earlier reviewed capture state 13:39:30.341 457414 −1.834322810 0.085667491
Later neutral state 13:39:30.475 457610 −1.901736736 0.018253565
Next fresh precontrol 13:39:46.966 490865 −1.919990301 0

Y/Z and yaw are unchanged at those boundaries; world20553048 and player29414172 remain the same, health/shield100. The original limit remains 0.08 world units and 4° yaw. The small neutral comparison is justified only with explicit review of the neutral image and matching state; it is not permission to substitute a later unreviewed state. I inspected the retained earlier, neutral and blocker PNGs: stone western lane, pistol and HUD100; no positively identified armed hostile. diagnosis.json retains full precision, image SHA256 and review notes.

The actual sequence is post memory → capture memory → registry read → PNG → short neutral memory → neutral PNG → human review delay → fresh precontrol. The earlier image was reviewed13:39:44.817313Z; the neutral image has no original explicit review entry. Exact historical PNG request/completion times and a post-neutral-PNG memory sample are absent. Memory precedes PNG, and the game continues during review. The neutral-to-fresh interval is16.491 host seconds; this does not identify physical key hold time, engine consumption time or an engine defect. Four original down/up delivery pairs remain the accepted input evidence; no fifth keydown or new guest delivery is claimed.

The helper creates a new memory → PNG → memory neutral pair with request boundary times and hashes, then requires explicit review naming the exact PNG and paired JSON digests. Both bracket states constrain the fresh precontrol sample. It rejects stale pairs, pending/unreviewed captures, superseded reviews, changed bytes, null states, damage/death, changed world/player identity and position/yaw drift. Brackets are bounded to1 host second, review age to60 seconds, and fresh precontrol age to1 second. These additional bounds do not make capture atomic. Approval permits one operation; the next needs a new pair and review. The adapter preserves release timers/finally release, post/neutral damage checks, input limits, original observer code and the visible-armed-target fire restriction. Rejected review/control stops without steering; no automatic rebaseline or retry.

Validation: 11 host unittest methods, including actual old rejection and meaningful negative controls; the byte-exact original four controller tests; and five unchanged source-only input/clock tests, including five clock mutation rejections, all pass. Tests replay retained records using synthetic clocks/transport; synthetic pair receipts are explicitly labeled. The original observer/control files and input/clock assertions are preserved and hashed. No production change or heavy build was needed for this documentation/evidence scope. Accepted input723e390a4 and coordinator-jedi-preflighted-review/review.json were reused without repeating packaging audits.

Executable helper/tests, selected original records, controller.patch, logs, fullsha.json, recovery-payload.tgz, RECOVERY.md, FINAL.md and result.json are retained in the assigned ignored run directory. No concrete engine runtime defect was found. No browser, boat, guest launch, deployment, predecessor lifecycle operation, cleanup of foreign files or subworker was used. No box was assigned; coordinator retains lifecycle ownership. Next: separately dispatched campaign using fresh reviewed neutral pairs, with ordinary first-blocker stopping.