Qualify route gain and observe native gate context through ordinary input

docs/prison-lock-alignment-plan.md

@@ -52,3 +52,18 @@ Use a separately explicit **alignment-only** variant after the independently ver
 The original input QUEUED clock is **30094ms**. Its unchanged expiry is **75094ms**, leaving approximately **11.3 seconds** after the observed approach completion. Roughly 42° of yaw, potentially 9.9° of pitch, native settling, and a four-second held action leave little margin. **Alignment-only is the next logical experiment.** Do not extend the observation or approach deadlines to fit F into this run.
 
 A later F-only proof must start with at least six seconds of the original observation budget remaining and stable exact selection, use one bounded F-down/up pair, retain the proposed five-second F lease without changing W’s lease, and prove native progress plus PRISON → RADIO_SEARCH and gate response while still inside. Until those prerequisites are observed, send no F. No runtime or input changes were made for this audit.
+
+## Implemented explicit yaw-only experiment
+
+`tools/headless_game.py --exercise-prison-alignment` requires the existing explicit approach, relative-look, native input/gate observations and private-display inspection. It runs only after the raw native approach has independently passed. The separate alignment policy receives no keyboard-sending capability. Up to six horizontal relative mouse requests, each at most128pixels and15degrees of observed turn, remain subject to current/projected pointer containment and the same private window/focus/prefix ownership checks. The cumulative limit is768pixels. Body drift is limited to0.05m; native movement, action progress and door changes refuse the stage.
+
+A separate immutable eight-second supervisor lease is shortened by the remaining original45-second native lifetime; neither approaching the gate nor starting alignment renews that lifetime. Native geometry, original actor/life/camera/lock/context, nonce, same-run signed calibration and ordered sample receipts remain bound. The stage requires two consecutive stable samples selecting the original lock and context with cached permission and genuinely bound zero progress. It never sends F, vertical movement, pitch or scroll input, and has no direct selection/action/world setter.
+
+The helper and independent raw-log validator are hashed separately from the approach modules. Default keyboard-only and approach-only modes remain explicit alternatives. Simulated checks establish policy and integration boundaries only; a completed native alignment archive is required before claiming that normal input selects the lock.
+
+
+## Native yaw experiment result,19September19:41
+
+The [194128 archived run](evidence/game-2026-09-19-194128-gate-context-unselected/runner.json.txt) remains failed: its independent [review](evidence/alignment-review-2026-09-19-194128/review.json) accepts the prior4.760m approach with10pulses and0.475m final waypoint distance. Four additional horizontal turns acquired the exact gate context, with cached lock permission true, but selected lock and bound progress false. The view had reached the requested yaw region; no F or door action was sent. The eight-second alignment supervisor stopped the run; all attempted W presses were released and private cleanup verified. Next is read-only selected-action/class/order diagnostics to distinguish the disabled stock door action from the available lock, before any separately bounded action-cycling experiment. No larger reach/radius or action-availability change is justified by this result.
+
+The earlier [193924 attempt](evidence/game-2026-09-19-193924-approach-stopping-reserve/runner.json.txt) stopped before alignment,0.982m from the interior waypoint, because the largest observed stop0.907m plus its0.25m margin exceeded remaining distance. It remains a failure; the fresh194128 attempt used unchanged limits. This also documents normal-input timing variability rather than establishing repeated approach reliability.