Skip to content

START_HERE.md

Latest commit

 

History

History
955 lines (874 loc) · 57.5 KB

START_HERE.md

File metadata and controls

955 lines (874 loc) · 57.5 KB
Raw

Touhou Solver Current Handoff

Last updated: 2026-08-27.

This is the only volatile entrypoint. Read AGENTS.md, GOAL.MD, then this file, STRATEGY.md, and the focused task card in notes/review/TH08_LUNATIC_NMNB_RESEARCH_TASKBOOK.md.

Historical material removed from the active tree is recoverable through ARCHIVE_INDEX.md and tag pre-workspace-prune-20260731. It has no current authority.

Active Scope Override

The immediate completion target is a fresh full-game Sakuya/Remilia Easy Route-2 Final-B NMNB. Easy is the infrastructure and local-policy proof before difficulty is raised again; the ultimate target remains Lunatic NMNB. The user authorized a separately cloned native Linux runtime as the search and replay-generation platform. The original Japanese v1.00d executable under isolated Wine remains the final authority: a Linux result counts only when its replay completes the same route there with zero native hit edges and zero Bomb input. Extra remains deferred, and fixed-root wind tunnels remain diagnostic.

The 2026-08-01 Hard/pause state is historical. Preserve its useful physical and authority evidence, but do not let its stale scope override the current user request. New source/runtime discrepancies are tracked in notes/review/TH08_SOURCE_AND_RUNTIME_AUDIT.md; the current architecture and ordered implementation plan are in notes/review/TH08_SOURCE_AUTHORITATIVE_SOLVER_AUDIT.md.

Checkpoint

  • Branch: codex/th08-linux-native-replay-bridge.
  • Pre-pivot checkpoint: main and annotated tag pre-linux-native-pivot-20260826 point to d7a66c5. The Linux runtime is a separate clone at source revision 4cffb2afa8d4a62083a5afc4a1968f51e96ac2cf on branch solver/linux-lockstep-replay-bridge; its verified i386 build is playable and exposes the original fixed target-owned globals. The adopted architecture pauses at the Linux DirectInput sampling boundary, excludes solver wait from the game-visible clock, and records a normal replay. It must preserve every original logical update, input, held-key, and RNG epoch. Linux NMNB is only a candidate; original-Wine replay NMNB is completion. The protocol, replay-identity caveat, numerical risks, and differential gate are specified in notes/architecture/TH08_LINUX_LOCKSTEP_REPLAY_AUTHORITY_20260826.md. The runtime branch is now at c38a92a; its verified ELF SHA-256 is 0bd9685c8c1a7cc57a4aaee9c1a449485b8539c9a6272bcb22d70c7a9fe9f1f1. The generic signed-RNG, opcode-148 clock, defeat-mode-2 effect-owner, and diagnostic retail-life corrections are active. The retained Stage-5 differential is exact for all 12,000 sampled roots through replay frame 12,494. Solver-side Linux replay generation and retail-life playback support are committed through 2d86b01. A normally saved Linux-origin Easy Stage-1 replay now terminates naturally in both a fresh Linux process and the unpatched canonical v1.00d executable. From manager/replay frame 100 through 2,162, all 2,063 aligned semantic roots are exact after excluding only sampler-address provenance. This closes the reverse-replay compatibility gate for the retained trajectory. A small synchronous one-epoch generic local-planner driver is now active: zero snapshot lag, zero pickup delay, one-epoch hold, complete current pools, and no historical Wine pending/no-write transaction model. A normal-start 3,600-epoch Easy smoke completed with zero hit/Bomb/input-echo mismatch and 337 non-stay focused/fast decisions. The next physical gate is one uncapped full Easy route; do not port the historical asynchronous Wine loop wholesale.
  • Current Easy physical/local checkpoint: accepted Final-B Practice easy_route2_stage6b_unattended_20260826_051733 remains the best result at nine hits and zero Bomb input, versus 18 hits in the paired baseline. Source and native state-5 evidence then closed the two apparent sensing gaps: existing bullet slots 1156 and 340 were under-advanced because the controller merged their pool clock with newer boss/ECL reads and negated an unrelated source offset. The implemented generic fix carries the bullet bracket's full discrete age support through local prefix, preflight, beam, delayed, and issue certificates. Exact pre-hit replay rejects the actually issued action at frames 11412 and 53730 while alternatives still exist. AUD-093/CE-0267 retain the proof. The fresh physical falsifier easy_route2_stage6b_unattended_20260826_063823 completed normally with 11 hits and zero Bomb input, so it is not a survival promotion. It did remove the residual observability class: all 11 contacts now have observed bullet/laser or modeled committed-prefix evidence, and ten reach robust action-set exhaustion before contact. AUD-094/CE-0268 retain this split. Mutually exclusive ages now reduce to one conservative source-AABB hull per native slot. Exact contacts, the 1,569-test suite, and a 360-check atomic/straddled stateful geometry gate pass; a 400-bullet local pass is about 13 percent faster than concatenation. The proposed unconditional 10..32-frame constant-action tail was then rejected offline: on eight deterministic complete stages it did not reduce aggregate hits, introduced two regressions, and paid substantial future/local latency. A same-future three-seed isolation still introduced one hit and added 10--23 percent median local latency. The conditional tail now uses the exact native binary32/time-scale movement recurrence, but remains dormant on ordinary Easy roots. AUD-094/CE-0269 retain this negative result. The next gate is an earlier-loss audit of the physical boundary/exhaustion roots, not a full route or a global-planner run. That audit first found that beam replay had substituted total issue age for player-to-hazard lag and omitted the discrete bullet-age support. Corrected replay now matches all eight sampled physical actions and fails closed on incomplete clock authority. Dense deadline roots still cost roughly 43--47 ms locally. Source audit then identified a simpler execution-profile error: the synchronous enemy prefix begins at native slot 1, while the legacy background worker makes hundreds of scalar process reads to supply slot 0 and the tail. The worker costs 91.40/134.38 ms median/p95 in the retained route even though serialized collision bodies never exceed slot 32. An attempted prefix-only shortcut was rejected before commit or play: EnemyCount counts slots encountered during the in-progress update scan and is not a current-occupancy oracle. The safe generic source-contiguous profile instead reads slots 0--63 synchronously and retains a complete asynchronous slots-0--479 scan as one contiguous 10.3 MiB process read with a reusable buffer. When ordinary global authority is disabled, Easy also selects the existing local-only controller. A rejected timeline-ABI experiment was exact but had no idle median gain; it was removed rather than retained as complexity. AUD-095--096 and CE-0270--0271 retain the evidence. The complete Linux suite passes 1,577 tests with five conditional skips. A 3,600-frame saturated source-stateful gate also passes 120/120 future joins, 360/360 geometry checks, and 1,414,390 lifecycle samples against the C oracle with zero lifecycle-position error.
  • Current offline source-stateful checkpoint: the replayable complete-stage laboratory still covers 480--12,000-frame programs, finite 1,536-bullet and 480-enemy pools, hard no-Bomb campaigns, source/C lockstep, geometry differential, and shrinking. The generic retained-root executor now also instantiates timeline and ECL 0x5A--0x5E child births through synchronous bootstrap and ascending native pool order. Direct fires and callbacks 12/14 share one ordered stream; callback 14 has an independent C transcription. Existing live-pool bullets receive exact bounded callback velocity/collision schedules only under a point-aligned three-clock certificate. Global lowering now shares the 21-type state-2/3/4 spawn lifecycle instead of treating preactivation bullets as ordinary lethal motion. CorridorSolution carries the callback-join version and runtime handle, and the controller can submit the composed pool under exact unit scale. New v2 retained roots also capture canonical active bullet/laser planner slots inside that same clock; a 319-bullet/two-laser stateful root round-trips to an identical H16 policy. The four existing physical Stage-5 roots are v1 and omit current hazards, so they cannot support the intended same-root solve. The same laboratory falsified the active-transform growth heuristic in 1,069/1,969 samples, worst 97.060802px error against a 31px bound. Global geometry v4 now makes every such current pool shadow-only. Composition is deferred until a global worker slot is due and free, and a worker-root certificate is not reused by later local prefixes. Current-hazard schema v2 now retains each complete 18-record program, queue/culling state, and all meaningful active handler blocks only on the coherent diagnostic path; packed local sensing is unchanged. Source/asset audit also proved that reflection/culling must use visual sprite geometry rather than the collision AABB; stateful schema v4 now carries both for all 21 real templates. The retained program now has a source-order scalar stepper and a separately compiled persistent C batch; long mixed histories agree, including queue order, shared blocks, wraps, movement, and culling. Template replacement and derived births fail closed. The 1,536-state native kernel takes 0.02446 ms median, but Python prepare/decode raises full wrapper time to 22.233 ms, so global integration must retain flat native state and cannot enter the 16 ms local loop. This is offline integration, not arbitrary ECL, a global same-root policy win, or physical hit authority. The next physical diagnostic is a complete Easy Sakuya/Remilia Route-2 hard-no-Bomb attempt; every Easy hit must be audited as infra/model/issue evidence. The ultimate target remains Lunatic NMNB. Then close template/derived births and connect the current-plus-future kernel to global shadow replay. Thirty-one focused transform/oracle tests and all 1,544 Linux tests pass with five conditional skips; the paper builds to nine pages. The retained compact 3,600-frame compatibility artifact remains artifacts/benchmarks/th08_source_stage_fuzzer_gate_20260825.json; routine successful raw reports are not retained.
  • Current Lunatic physical checkpoint: lunatic_route2_fullrun_unattended_20260824_051944 completed Sakuya/Remilia Route 2 through Final-B at frame 230561 with 60 native hit edges, stage counts 1/4/8/15/14/18, zero Bomb input, and exact Wine-prefix cleanup. It physically validates OPT-002A: ten issue-deadline misses all recovered, none is on a hit's causal row, and the former 32-decision Stage-4A self-lock did not recur. It does not validate a hit-rate improvement: OPT-001 had 61 hits on a different RNG root. Fifty-seven of 60 residual hits follow robust action-set exhaustion, and the global corridor recorded zero submissions or queries across all 57,553 decisions. Do not run another route or change focus/boundary scoring before reading notes/review/TH08_SOURCE_AUTHORITATIVE_SOLVER_AUDIT.md.
  • Historical Hard checkpoint: hard_route2_stage4a_unattended_20260801_191508. It completed Hard Stage 4A with 18 hits, first hit 1975, zero Bombs, accepted replay, and complete cleanup. Twelve hits were boundary-associated and eleven used fast movement. This is a regression/falsifier, not a promoted outcome.
  • Observed no-scale-writer mechanism pass: Stage-4A 184903 bound the exact runtime/static ECL identity, coherent unit root, stage/difficulty/ route context, and complete installed-callback inventory. It restored 3,599 constrained decisions from zero in 183730; spell-56/60/64/68/72 received 872/547/164/566/529 constraints, and recovery/distant-recovery was selected 319/809 times. The run completed 12 hits, first hit 1209, and zero Bombs. Different-RNG outcome counts remain observational; action-authority activation is observed.
  • Observed hard-authority seam: 184903 also relaxed 57 exact-corridor constraints through the coarse terminal-threat fallback. Exact corridor, ordinary predecessor, held, delayed, and continuation-lease authorities are now non-relaxable. 191508 recorded zero such relaxations.
  • Observed continuous-position falsifier: the old 16px global corridor represented only cell centers with zero sampling-radius clearance. The current lower bound consumes the complete cell half-diagonal sqrt(2)*8 = 11.3137px. On 191508, queryable rows increased 5,104 to 5,775 and median solve time fell 131.5 to 78.7 ms, but empty action sets increased 1,723 to 2,927, constrained decisions fell 3,599 to 3,001, and hits rose 12 to 18 on a different RNG root. The semantic hardening is retained; a uniform 16px lower kernel is physically ineffective and must not be called a solution.
  • Observed Hard regression: historical checkpoint 709e858 completed a different-RNG Hard Stage 4A trial 20260726_212756 with six hits, one nonspell hit, 13,203 queryable global-policy rows, 8,073 constrained decisions, 725 selected recovery rows, 2,845 selected distant-recovery rows, decision cadence median/p95 2/3 frames, and local planning median/p95/max 11.686/20.830/38.136 ms. Current 175112 had only 4,737 queryable rows, 847 constrained decisions, zero selected recovery/distant recovery rows, cadence 3/4, and local planning 17.953/32.725/250.616 ms. Different-RNG hit totals are observational; the missing authority, zero selections, and latency tails are direct repeated delivery mechanisms.
  • Observed primary cause: all 4,150 current spell policy queries had zero constrained decisions. The solver computed viable spell actions, but the schedule provenance was diagnostic_constant_current_root_unknown_direction_no_authority, so time_scale.hard_authority and corridor action_authority were false. Later SEM-SCALE work correctly revoked the old unsound constant-root assumption, but no sound Stage 1–5 replacement was installed. Work then concentrated on the sparse ordinary exact lane while most Hard hits were in spells.
  • Observed/inferred replacement boundary: shipped native dataflow has gameplay scale writes in ECL callbacks 18/28/29; other writes reset scale at game/stage initialization. Exact decoded ECL for Stages 1–5 contains no literal or dynamic invoke/install of callbacks 18/28/29. Stage 6 and Final do contain callback 18 and therefore cannot use the same certificate. The next implementation is a runtime-ECL-identity-bound, unit-root, no-scale-writer schedule authority that fails closed on content, context, root, or callback-inventory mismatch.
  • Observed source-v15 physical result: 519/1,065 source roots completed, 436 were truncated, and 83 carried the full H268. The v15-targeted dynamic type/color, 0x19, 0x25, 0x35, 0x6F, and supported-transform failure reasons disappeared. Exact prepublication/delayed/lease authority was effective on 229/17/16 issues, with ten lease creates and seven renewals. Source v15 therefore passes its named semantic gate but does not improve the different-RNG hit total.
  • Observed canonical computation-gap falsifier: the first hit's prior 240-frame window contained 13 full-H268 source roots and exact prepublication authority on 43/50 decisions. Thus source publication was no longer absent before this hit. At the last safe issue f565, down_left was certified only through terminal f571 and no continuation lease was created. The next f566 capture launched a 394.522 ms synchronous delayed scan; input was not revisited until f594, 28 physical frames later. Retained bullet slot 1 crosses the held down_left path at f587 (signed clearance -0.432), so the generated sensor_gap label is false: the bullet was observed, but the finite terminal proof expired during computation. The next correction is a hard held-action computation guard plus viable-kernel-informed exact probe ordering, not another ECL opcode.
  • Implemented computation guard, awaiting physical falsifier: the former held-only H80 exact certificate is now a prerequisite for a synchronous delayed scan whenever no compatible lease exists. It consumes the complete native bullet/laser/body slab, future-source projection, no-write semantics, and the full effective issue-age/pickup support. An unsafe or unavailable held certificate blocks the scan and returns immediately to the bounded issue path; it grants no replacement action. The three-action terminal probe now orders current-kernel repair-volume candidates before fixed compass defaults, but each candidate still requires its independent exact prefix and terminal membership. The retained f566 slot-1 certificate is losing and therefore blocks the former 394.522 ms scan. Focused Linux and Windows gates pass 33/33. Rotate the physical falsifier to Stage 4A; do not infer success from aggregate hits alone.
  • Observed source-v14 physical result: 801/1,275 source roots completed versus 645/1,235 in source-v13 155718; 58 versus 47 carried the full H268. The targeted main 0x05/0x06 and captured movement-state-2 failures disappeared. Exact prepublication/delayed/lease authority was effective on 316/7/5 issues, with three lease creates and four renewals. This physically validates the v14 semantic correction, not route improvement.
  • Observed remaining latency/coverage mismatch: all three ordinary hits still had zero exact authority in their prior 240-frame windows. Before canonical hit f1253, the first full-H268 root was f1223 and was submitted at f1230, but its 2352.884 ms solve completed at f1371, after the hit and after expiry. Closing reached main/auxiliary semantics must move complete coverage more than the observed roughly 141-frame publication delay. The next concrete blockers are main float-add 0x19, main/auxiliary transform definition 0x6F, auxiliary return 0x35, dynamic type/color, auxiliary 0x25, and the resulting active transform programs.
  • Implemented source semantics v15; physical semantic gate passed: shipped code confirms direct-fire type/color are independent dynamic i16 operands, 0x25 normalizes one float lvalue, depth-zero auxiliary 0x35 terminates without a saved-frame restore, and 0x6F writes one exact 24-byte transform record. Reached main 0x19/0x1A now share the bounded auxiliary arithmetic. Active vector/angular acceleration, re-aim, and reflection use a conservative full-disc path bound; template replacement consumes the maximum reached template hitbox. Unsupported stack state, set-valued transform fields, and other transform kinds remain UNKNOWN. A real Stage-3 time-65 wave now closes through six fires, including active angular-velocity geometry. 93 Linux and 93 Windows focused tests pass. Physical 171249 removed the named failure reasons and delivered exact authority in the canonical pre-hit window; its new blocker is unsafe synchronous computation after a finite terminal certificate, not missing source lead.
  • Observed lease-v4 physical result: one lease was created and one renewed; 31 decisions captured it, 30 selected it, and 29 exact no-write issues were effective from f1725 through f1783. All retained the certified down_left_fast movement, emitted no physical input transition, passed the contact-body seam, and ended only at terminal expiry. Lease v4 therefore passes its narrow persistence/mechanism gate. It does not pass the global outcome gate: the run had five ordinary hits, and every ordinary hit's prior 240-frame window had zero effective prepublication, delayed, or lease authority.
  • Observed next source blocker: the common pre-hit status was future_policy_unavailable. Source roots frequently reported a complete causal prefix but were truncated below H268 by reached main ECL opcode 0x05/0x06 or captured hostile movement state 2. Before canonical hit f1263, the first nontruncated full-H268 root appeared only at f1249, too late to solve and publish. Exact closure of these reached source states is now the next general gate; local ranking and stage waypoints remain behind it.
  • Implemented source semantics v14; physical semantic gate passed: shipped enemy_ecl_vm_step confirms main opcodes 0x05 and 0x06 share the already-bounded local loop/set semantics. Shipped enemy_motion_update confirms state 2 requires stored displacement +0x2DC4, start +0x2DD0, easing bits 14..16, Th08Timer fraction/current at +0x2DE0/+0x2DE4, and duration +0x2DE8. Native projection v14 captures those fields and corrects the former +0x2DDC previous/current timer confusion, which also affected state 3. Missing or malformed timed state remains UNKNOWN. Linux focused and 147 Windows tests pass. Old retained state-2/3 roots cannot be uplifted. Physical 163434 confirms more complete and full-H268 roots, but not enough pre-hit lead for publication.
  • Observed Stage-3 first mismatch: 701/1,214 source attempts failed because timeline 2 retained a nonzero fractional timer component; 465 crossed the native clock bracket. Every one of the 49 source attempts in the f7695..7935 pressure window failed on the fraction, leaving no global submission before the hit. Native unit-scale timer advance at 0x00447421 preserves that fraction while incrementing the integer clock, and timeline instruction eligibility observes only the integer component. Source semantics v12 now validates finite timer state and discards the causally inert fraction only after exact unit scale is proved. Nonunit scale and nonfinite state remain fail closed. Focused Linux and 88 Windows tests pass.
  • Observed physical activation and remaining discontinuity: the corrected run produced 692/1,261 complete source projections, 596 submissions, 595 completed policies, and 265 effective ordinary exact-authority issues. Four delayed predecessors were effective and created four leases, but no lease was consumed: every lease was revoked by an uncontained fresh-body envelope. Before first hit f3181, exact authority ended at f1406; all 103 decisions in f2941..3181 were future_policy_unavailable despite 14 complete source captures in that window. The fraction correction is physically validated as source delivery, not as hit reduction or persistent recursive viability. Parent checkpoint 5fc6d92 then completed Stage 4A with 18 hits/10 nonspell versus its direct 21/12 baseline, but first hit was nearly unchanged at 833 versus 819; Stage 5 repeated exactly 18/10 versus its direct baseline. Both are different-RNG observations. Across Stage 3/4A/5, every created lease remained ineffective and fresh-geometry revocation dominated. The next gate is therefore the common active-hostile trajectory/lease recertification seam, not another aggregate rerun.
  • Observed physical latency correction: the ordinary fallback now builds an action-conditioned H80 predecessor indexed by final observed issue age, including computation, old pending evolution, held no-write, replacement, pickup delay 0..6, complete future births, and a fresh hostile-body slab. Packed player paths are bitwise equal to native-order scalar movement. Incremental action evaluation is ordered by a no-authority local proposal and reuses that immutable proposal after the scan. In 115838, 11/21 scans physically acquired exact authority; median issue age was 25 frames versus 43 before proposal reuse. CE-0250's zero-transition blocker is closed.
  • Implemented after cross-stage physical falsification, awaiting v13 physical rotation: an effective delayed action creates an immutable H80 terminal-continuation lease. Each compatible fallback root first renews the held no-write action; only if that predecessor is unsafe may independent recovery directions be evaluated. Exact float32 position, active/held/pending support, epoch, stage, phase, unit scale, remaining horizon, fresh geometry, and the full content-addressed projection version are fail-closed. Spell-active and stage-route are reread at final issue, so computation may not carry an ordinary lease across a phase/context transition. A new complete mask requires a new exact delayed predecessor. Stage-4A f3191→f3192 exposed a native phase bug: priority-17 current-input publication precedes its first priority-9 movement by one physical frame. Lease v3 retains that input-phase correction and requires source semantics v13, which adds every contact-enabled active hostile body to the exact ECL/native-motion trajectory from the projection root. Fresh geometry now conditions only the native observation seam at its own frame; it cannot replace the old causal proof with a restarted linear H80 future. Unstable observations, uncontained current bodies, older source semantics, and incomplete topology still fail closed. Rotated Stage-3 physical validation is next; scalar reserve remains rejected and local ranking is not safety authority.
  • Native H=32 wind-tunnel checkpoint: 3d15953; it is historical evidence, not the live ordinary horizon.
  • Workspace-prune checkpoint: be3e583 (Prune TH08 active research workspace). It removed dormant supplemental, candidate, prewarm, G5, priority-17, and old focused Final-B lanes while preserving the promoted baseline/pre-loss live path, native snapshot executor, exact pipeline workspace, and Final-B scale authority.
  • Affected Linux focused gates and 186 Windows focused tests pass, including input-phase, continuation-lease geometry, enemy-mode, issue-stage, preexhaustion, prepublication, controller, and retained-source tests. Complete Linux discovery ran 1,284 tests in 10.659 seconds and exposed five unrelated workspace failures: four pre-existing native ABI header/binary-versus- manifest mismatches and the preserved user-edited factorized report's worker count 8 versus the old test expectation 16. Do not rewrite that user artifact or rebuild unrelated native outputs as part of this checkpoint.
  • Implemented, factorized hard lower authority: default-off --ordinary-preexhaustion-authority no longer uses scalar reserve or the coarse signed terminal tensor. Ordinary nonspells build a 4px Boolean viability kernel with the 2.828px cell radius consumed as required clearance. At every issue, the active/held/pending predecessor holds the selected complete mask with no further writes across every pickup branch into the next active-policy layer. A still-pending command must be a member of that layer's exact action set.
  • Implemented, deterministic hard gate passed: collision-control projection v14 now captures the manager singleton and every active ordinary ECL/timeline source, auxiliary VM, installed callback gate, emission descriptor, phase state, and reconstructable motion state. Legacy timed roots are not uplifted. Reachable timeline spawns and ECL emissions are lowered fail-closed into consumed annular-sector bullet and hostile-body AABB trajectories for 268 frames. Unsupported opcodes, callbacks, timeline gates, motion, content identity, spell state, or nonunit scale remain UNKNOWN.
  • The exact future projection is retained with the asynchronously published policy. Its geometry is now consumed twice under one version: every physical frame in the active/held/pending root-to-publication certificate, then every frame of the future corridor policy. Coverage metadata alone cannot authorize an action.
  • In the physical gate, all 7,202 nonspell decisions were incorrectly rejected by the player_transition_or_predeath eligibility check, so the filter and early-kill preference each affected zero decisions. Native evidence shows player +0xE2A68 retains a deathbomb-window limit and is not zero while alive. More importantly, an offline counterfactual with only that check removed still permits the canonical down_left action and degenerates to all 17 actions under an uncontrollable prefix and at a clamped boundary. Scalar boundary reserve is not equivalent to global viability; do not rerun this design.
  • Observed correction: forcing the real 80-frame horizon makes the former 16px signed predecessor empty at all five retained roots. Its earlier eight-action f817 result used H=32 and is not live authority.
  • Observed deterministic replacement: the same immutable f817/833/835/850/910 chain has complete source/version coverage under the 4px cell-radius lower kernel. Native exact recurrence and the actual live adapter independently agree that f817 permits only left_fast/down_left_fast, excluding active up_fast and issued up_left_fast. f833/f835/f850/f910 are already empty and remain fail closed. Windows measured f817 kernel/certificate at 1412/21 ms; the five roots have no unresolved action. This is finite-model evidence, not a physical outcome.
  • Observed physical integration falsifier: Stage-4A run 20260731_220830 completed with 23 hits, first hit 939, and zero Bombs. Its different-RNG aggregate does not falsify the f817 directional result. Decisively, none of 1,754 live future-source projections completed: 1,726 crossed the manager-frame bracket. All 10,016 ordinary-authority records therefore failed closed and zero actions were constrained. This falsified live source capture/delivery, not the finite-model predecessor.
  • Implemented after that run: manager singleton plus all 480 ordinary slots are now captured in one contiguous 9.8 MiB read instead of at least 481 sparse process reads. Source semantics v3 also lowers the reached dynamic direct-fire count, auxiliary subtraction, indexed-enemy +0x2D30, and set-valued FRScreen gate cases. Reached unsupported events still fail closed.
  • Implemented exact delivery acceleration: the Boolean native kernel now rejects branches by next-layer terminal membership before walking hazard samples and publishes an exact all-action kernel when the whole clearance slab/layer proves every clamped transition safe. Hard membership no longer computes soft repair/recovery scans eagerly; directional recovery remains available when the Boolean predecessor is empty. Windows no-hazard viability fell from the physical 2,996 ms bootstrap solve to 282 ms cold and about 9 ms warm. The retained dense-root worst case is still 2,104 ms and remains fail-closed on expiry.
  • Observed contiguous-capture falsifier: correctly configured Stage-4A run 20260731_231944 completed with 35 hits, first hit 498, and zero Bombs, but only 1/1,859 source projections completed; 1,836 crossed the manager frame and zero decisions received hard authority. The earlier 20260731_230657 replay omitted the runtime ECL identity flags and is an invalid authority experiment, not an 18-hit improvement.
  • Implemented after that run: the dedicated source worker now allocates its 10.32 MiB destination before the frame bracket, reads directly into it, and exposes manager/pool memoryviews. This removes about 10.6 ms of Windows allocation/.raw/slice copies without relaxing exact frame equality or future-source coverage.
  • Observed physical authority activation: Stage-4A run 20260731_233856 completed with 24 hits, first hit 845, zero Bombs, an accepted replay, and full cleanup. Stable capture produced 119/1,914 complete roots; 196 ordinary decisions were constrained and 193 remained effective at issue. This proves the exact authority is physically live. It does not promote the outcome: all 98 pre-first-hit applicable roots were early all-17 safe sets, while the pressure wave had 295 unavailable roots. No nonempty effective authority appeared within 80 frames of a hit.
  • Observed semantic starvation: the dominant complete-root blockers were 699 manager-frame crossings, 476 legal auxiliary timer roots, 237 dynamic float variable 10069, and 181 armed phase successors. The first two semantic classes plus generic auxiliary opcode-0x02 wait scheduling are now lowered from shipped-code evidence in source semantics v4/projection v13. Armed phase successor coverage remains fail closed.
  • Observed source-v4 delivery falsifier and implemented correction: Stage-4A 20260801_002006 completed 24/1205/no-Bomb. It produced 159/1,942 complete projections, but exact authority was effective on 135/6,146 nonspell decisions and 0/383 decisions in the 80-frame nonspell hit windows. Complete roots 1137..1200 were stranded behind an UNKNOWN-source future publication. Ordinary mode now spends no solver slot on incomplete source coverage and exposes a completed exact pending policy to the causal pre-publication predecessor before its epoch. The next gate tests this delivery fix; phase/transform/opcode/callback/movement closure remains open.
  • Observed pending-delivery activation and next correction: Stage-4A 20260801_004533 completed 22/1837/no-Bomb. Exact ordinary authority was effective on 287/7,275 nonspell decisions; 110 effective roots used a completed pending policy and 43 did so before the first hit. This validates the CE-0244 delivery correction. Authority was nevertheless effective on 0/287 decisions in the 80-frame windows before 11 nonspell hits. Complete sources carried up to 750 future sectors, and representative 6.9/19.8/23.5 second solves spent 6.1/19.5/23.2 seconds building clearance but only 0.81/0.24/0.23 seconds in viability.
  • Implemented exact sector-volume acceleration: native geometry now groups identical origin/radial/inflation samples, evaluates their exact circular angular union once per lattice point, and parallelizes disjoint frame slabs over at most 16 workers. A synthetic 750-trajectory, 81-frame, 4px regression completed in about 198 ms and matched the independent Python signed-clearance oracle exactly at sampled frames; 80 randomized angular unions also had zero observed error. Linux/Windows focused native parity and the retained f817/833/835/850/910 semantic chain pass.
  • Observed sector win and newly isolated recurrence deadline: Stage-4A 20260801_011902 completed 26/796/no-Bomb with an accepted replay and full cleanup. Different-RNG hit count is observational. Exact sector clearance fell from the preceding physical maximum 23.2 seconds to 0.993 seconds, while authority was effective on 399/7,162 nonspell decisions. It was still effective on 0/380 decisions in the 80-frame nonspell hit windows. At the canonical hit, source 639 spent 478 ms in clearance and 918 ms in Boolean viability; it arrived after the corridor's last nonnegative gate at frame 714 and the predecessor was empty by frame 716. Future-source failures, especially 1,195 armed phase transitions, dominate later aggregate gaps but do not explain this first hit.
  • Implemented exact interior recurrence acceleration: each Boolean layer now proves an all-action state in one set-valued reachable box only when every covered signed-clearance cell and every terminal action is safe. All unproved states use the unchanged branch recurrence. On Windows the retained f817/833/835/850/910 chain keeps exactly the same action sets and no unresolved action while maximum solve time falls from 2,071 to 775 ms. A fresh Stage-4A physical gate is next; Stage 5 remains withheld.
  • Observed recurrence win and coverage handoff: Stage-4A 20260801_015631 completed 20/2009/no-Bomb with accepted replay and cleanup. Different-RNG hits are observational. Nonspell exact solve maximum fell from 3,050 to 1,338 ms and viability maximum/p95 from 3,026/1,664 to 1,038/926 ms. Before the first hit, 278 decisions had exact applicable authority and 102 of those were directional, but authority was still 0/238 in the 80-frame windows before nine nonspell hits. Every first-hit-window decision was future_policy_unavailable; roots 1915..2007 chiefly failed on nonzero transform programs, with opcode-0x19 interleaved.
  • Implemented bounded transform/addition closure: shipped-code recheck confirms only transform records whose kind intersects the bullet's original flags execute. Active 0x80 stop/re-aim is lowered as a full-direction path-length disc using the maximum emitted/resume speed, without predicting the future player. Barrier, culling-suppression, and sound records are movement-neutral; every other active kind remains UNKNOWN. Auxiliary opcode-0x19 now performs exact interval addition. The retained Windows chain keeps identical action sets/no unresolved action. Physical revalidation is next; Stage 5 remains withheld.
  • Observed transform closure and issue-time miss: Stage-4A 20260801_022228 completed 19/1833/no-Bomb with accepted replay and cleanup. Transform and opcode-0x19 projection failures disappeared. For the first time, three decisions in a nonspell 80-frame hit window had directional exact sets of 5/8/6 actions. All three missed issue deadline, however, so effective pressure-window authority remained zero. At frames 3501/3518/3532 the held up_right input was outside or not selected from the exact set, but 12..16 frames elapsed after capture against pickup-support high 6 and the deadline guard correctly held the old input.
  • Implemented causal delivery QoS and next reached flow: ordinary exact solves now run below-normal on their background owner and use eight native workers, leaving capacity for TH08/sensing/issue control. The Windows retained-chain maximum rises from 804 to 1,095 ms but remains within the 80-frame initial lead; exact action sets and unresolved status are unchanged. Reached auxiliary 0x05 loop-jump and 0x2E integer-LE jump are also advanced exactly; other flow remains fail closed. Revalidate Stage 4A; Stage 5 remains withheld until effective pressure-window authority appears.
  • Observed QoS falsifier and bounded phase correction: Stage-4A 20260801_024419 completed 22/1291/no-Bomb with accepted replay and cleanup. Eight-worker/background-low-priority execution was active; solve p95 changed 230→218 ms, but a 5,290-ms tail remained and hit-window effective directional authority stayed zero. 0x05/0x2E failures disappeared; ordinary slot-0 armed phase transition became 1,243/1,714 incomplete roots. Source v6 captures health successors and the exact phase timer and ignores only a timeout proved beyond elapsed + horizon; health or reachable timeout transitions still fail closed. Linux/Windows tests and the retained exact action sets pass. Revalidate Stage 4A before any Stage-5 run.
  • --kill-before-saturation now uses observed ordinary bodies only. The falsified timeline spawn forecast is withheld from live input. Observed-body alignment/unfocus remains a proposed objective, but the rejected reserve set is no longer an eligibility source; exact ordinary viable membership and fresh issue safety are required before another live gate.
  • Linux route bootstrap is live: commit 4aa6e69 reads source-defined title state and emits press/release edges to select Start, Easy, and Sakuya/Remilia. Two isolated native launches reached completed Stage-1 loading with difficulty 0 and shot type 2 at epochs 140 and 138. This is menu/lifecycle authority only; replay and numerical differentials remain open.
  • Dynamic bridge evidence corrected: commit 5ab79d2 distinguishes supervisor input 0x164d528/0x164d530 from adjacent GUI replay input and matches the true wire globals plus RNG through a 134-epoch live route bootstrap. It also retries the real Unix connect across the bind/listen startup gap; no route-duration timeout was added.
  • audits/ and archive/ are untracked/local. Never stage them.

Current Outcome

Physical baselines

Latest user-authorized Lunatic Route-2 practice ring:

Workload Run Hits First hit Bombs Replay
Stage 3 lease-v4 persistence gate 20260801_155718 7 1263 0 accepted
Stage 3 source-v14 delivery gate 20260801_163434 4 1253 0 accepted
Stage 3 20260731_091104 5 2150 0 accepted
Stage 4A 20260731_091925 13 2555 0 accepted
Stage 5 20260731_093027 12 2124 0 accepted
Stage 5 early-kill gate 20260731_122855 13 6981 0 accepted
Stage 4A global/early-kill gate 20260731_130103 16 1827 0 accepted
Stage 4A pre-exhaustion early-kill gate 20260731_133852 11 4148 0 accepted
Stage 4A forecast/global investigation 20260731_142342 18 1915 0 accepted
Stage 4A reserve-authority falsifier 20260731_152921 17 914 0 accepted
Stage 4A live-source/delivery falsifier 20260731_220830 23 939 0 accepted
Stage 4A ECL-config-invalid replay 20260731_230657 18 499 0 invalid gate
Stage 4A contiguous-capture falsifier 20260731_231944 35 498 0 accepted
Stage 4A exact-authority activation 20260731_233856 24 845 0 accepted
Stage 4A source-v4/delivery falsifier 20260801_002006 24 1205 0 accepted
Stage 4A pending-delivery/clearance falsifier 20260801_004533 22 1837 0 accepted
Stage 4A sector/viability deadline falsifier 20260801_011902 26 796 0 accepted
Stage 4A recurrence/coverage handoff 20260801_015631 20 2009 0 accepted
Stage 4A transform/delivery falsifier 20260801_022228 19 1833 0 accepted
Stage 4A delivery/phase falsifier 20260801_024419 22 1291 0 accepted

The automatic older-root comparisons were 15→5, 10→13, and 19→12. They are observational only: RNG roots differ and the proposed WS-H strategies were disabled in the original ring. The later early-kill gate is also different-RNG: it physically applied 27 certified unfocus preferences, delayed first hit by 4857 frames relative to the listed Stage-5 baseline, but worsened total hits by one.

Stage-4A run 20260731_152921 is different-RNG and its 17/914 aggregate is observational. Its experiment activation is decisive: the flag was present on all 12,029 decisions but yielded zero eligible, applicable, or effective constraints and zero early-kill applications. All 7,202 nonspell decisions failed on a stale native-field interpretation. The canonical hit also falsifies a gate-only repair. Contact bullet slot 455 first entered the retained nearby set at decision 817; the snapshot-801 global policy still certified down_left through frame 833, while the snapshot-818 policy delivered an empty set at 835. The compact trace cannot attribute that version flip to slot 455 alone. Counterfactual reserve evaluation still allowed down_left there and allowed all 17 actions at frames 850 and 910. Robust local prefixes exhausted only at 910, four frames before the hit.

The newer Stage-4A run 20260731_220830 is also different-RNG; its 23/939 aggregate is observational and must not erase the earlier local improvement. Its exact failure is live integration: all 10,016 authority records had incomplete coverage and no allowed set. Of 1,754 future-source attempts, zero completed; 1,726 crossed enemy_manager_frame, 16 observed impossible auxiliary depth from a torn root, four lacked the then-uncaptured indexed timeline field, three reached a dynamic direct-fire count, two hit an FRScreen message-clock boundary, and three reached the now-lowered auxiliary 0x1A. The worker also produced 3,610 expired-policy and 3,792 no-query decisions. These are capture/publication counterexamples, not evidence that left_fast/down_left_fast at retained f817 was wrong.

Stage-4A run 20260731_231944 supplied the correctly configured follow-up. Only 1/1,859 projections completed and 1,836 crossed manager frames; all 4,499 state-eligible nonspell decisions remained future_policy_unavailable, including all 332 before first hit 498. Its 35 hits are fallback-path evidence. The preceding 20260731_230657 run had no runtime ECL image, emitted zero source projections, and is retained only as a configuration counterexample.

Stage-4A run 20260731_233856 closes the capture/activation question: 119/1,914 source projections completed, 205 decisions were authority-eligible, 196 were constrained, and 193 were effective at issue. There were 24 nontrivial directional sets and nine exact empty predecessors; the early applicable sets before hit 845 were all 17-action safe sets. The 24-hit different-RNG aggregate is not a causal regression verdict. Its actionable falsifier is pressure-wave semantic starvation, not authority wiring.

Stage-4A 20260801_002006 verifies that source v4 removed the earlier 10069 and legal auxiliary-delay-root failures, but it did not deliver useful authority near hits. Exact authority was effective on none of 383 decisions inside the 80-frame windows before the 14 nonspell hits. At the first hit, complete roots 1137..1200 existed while an UNKNOWN-source future publication blocked the serial worker. This is now corrected by gating solve submission on complete source coverage and treating a completed exact pending policy as the pre-publication terminal kernel. Stage 5 was correctly skipped.

Stage-4A 20260801_004533 physically validates that correction: 110 effective ordinary decisions used the exact pending terminal. It also isolates the next deadline blocker. Dense complete future sources reached 750 sector trajectories and spent up to 23.2 seconds in clearance-volume construction, leaving zero effective authority in all nonspell hit windows. Native grouped angular-union/frame-parallel clearance is now deterministic-parity validated; its physical effect has not yet been measured.

The physical forecast gate also falsified the current timeline observer as a general later-wave source. All 376 observations recycled the same timeline-0 startup birth at time 1, x=30, and only three affected input. Full-health observed-body targeting remains useful, but zero-pointer timeline lifecycle must be disambiguated before the birth forecast is trusted again. The live forecaster is now disabled rather than used as an eligibility source.

Latest full game-start Lunatic Route-2 run:

  • lunatic_route2_fullrun_unattended_20260824_051944
  • 60 hits, zero Bomb inputs
  • stage counts 1/4/8/15/14/18
  • reached route_complete
  • physically validates OPT-002A deadline feedback and preserves OPT-001/002 sensing improvements;
  • 57/60 hits follow robust action-set exhaustion; causes are 41 modeled committed-prefix collisions, 16 observed bullets, two observed lasers, and one enemy body;
  • 48 boundary and 43 fast-mode labels are correlational. Source audit found a 2px planner radius versus the physical 1px half-extents, nonlethal bullet states admitted as hazards, and capsule-versus-rectangle laser mismatch;
  • all 57,553 corridor submissions were hard-authority blocked, so worker and global query counts remained zero. The VPS did no strategic search.

The preceding OPT-001 and OPT-002 routes remain useful mechanism references: ...022909 had 61 hits and ...034510 had 67. Their different native RNG and post-hit cascades make all hit deltas observational.

The historical Windows run lunatic_route2_fullrun_unattended_20260730_222529 had 68 hits with stage counts 2/3/5/20/15/23, but used pretarget guidance that is not accepted as source-authoritative under the current contract. Treat its lower latency and guidance as diagnostic rather than promotion authority.

Retained dossiers are under notes/runs/. Compact reports and valid practice replays remain under artifacts/.

Native wind tunnel

Canonical Stage-5 replay first hit:

  • manager frame 2136;
  • root frame 2129;
  • recorded mask 0x05;
  • hostile bullet slot 45;
  • signed separation -0.966766.

The rolling executor replaces only the fixed native calculation-chain call, holds the owner at the root, freezes unrelated threads, executes original TH08 update code, and restores a verified same-session root. Parent replay is bit/exact-state checked before branch authority.

Observed result:

  • all 36 no-Bomb root masks were executed in original TH08;
  • 324 causal branches were searched;
  • a warm session executed 180 branches in 309.089 seconds with exact parent repeats;
  • policy 0x94 -> 0x44 -> 0x10 -> 0xA4 at frames 2129/2137/2145/2153 stays unhit through frame 2161;
  • an H=32 natural frame pump matches 32/32 native snapshot ticks.

This proves one exact fixed-root, fixed-horizon original-engine witness. It does not prove a full spell, live delivery, or physical NMNB.

Primary evidence:

  • artifacts/runtime_reports/th08_native_snapshot_causal_policy_root2129_h32_20260730.json
  • artifacts/runtime_reports/th08_native_model_trajectory_root2129_h32_20260730.json
  • artifacts/runtime_reports/th08_native_model_consumable_h1_root2129_20260730.json
  • artifacts/runtime_reports/th08_native_state2_lifecycle_root2129_h8_20260730.json
  • artifacts/runtime_reports/th08_native_h1_ecl_source_differential_root2129_20260730.json
  • notes/architecture/NATIVE_REPLAY_CAUSAL_WIND_TUNNEL_AND_REPLAY_SAVE_CONTRACT_20260730.md

The old closed-form slot-45 forecast first differed by one x ULP. A corrected per-update binary32 recurrence matches the retained native fixture. The current source differential still returns UNKNOWN at the unresolved pre-enemy/pre-aux producer; do not infer it from retrospective RNG alignment.

One mapping-epoch poison event is retained. A future persistent warm service must use single-writer ownership, immutable session/root IDs, branch validation, cooperative cancellation, idle TTL, poison cleanup, and automatic rebootstrap. Do not weaken the epoch gate for speed.

Combat/resource model

The active WS-H reconstruction now covers Route-2 normal shots, supported native damage, enemy generations, defeat/cleanup distinction, Boss transition identity, item allocation/pickup, Power/resources, and mandatory timeline events. Its general combat model remains offline. The only live combat experiment is the narrow default-off ordinary-enemy alignment preference described above. It has no independent safety authority.

The immediate high-value hypothesis is not another schema:

  • spells: survival first;
  • ordinary enemies: inside the survival-feasible set, test whether earlier kills prevent later saturation;
  • dynamically compare focused micro-control with unfocused fast movement and shot coverage;
  • collect early Power only when a safe path produces a causal later benefit.

The first Stage-5 root-4300 test is complete. Observed: clearing Focus for eight ticks defeated one 20-HP enemy, suppressed three later hostile bullets per nine-frame emission cadence, and reduced the endpoint bullet count 548→539. Observed: the current offline global layer-0 viable masks and safe-action masks remained exact-equal at frames 4314/4323/4332. That offline result did not end the hypothesis; the physical gate below exercised the rule 27 times, but global live guidance still never published.

See notes/CURRENT_COMBAT_RESOURCE_MODEL.md.

Live Authority

The physical policy is hard no-Bomb:

  1. native state sensing and hazard projection;
  2. native packed bullet decode;
  3. robust Boolean viability;
  4. baseline local beam plus promoted pre-loss continuation preference;
  5. fresh issue-time collision certificate;
  6. exact-version action transaction and fail-safe fallback.

The latest full route loads the pinned Final-B ECL image, but its transported unit schedule remains experimental_pretarget_unit_transport_unknown_direction. Under authority-only corridor policy this is not a hard schedule: every due submission was blocked and no global policy constrained input. The root-only constant-scale continuation remains diagnostic and unknown-direction, never a general hard-safety authority.

The rejected scalar pre-exhaustion experiment is no longer connected to live input. The replacement remains default-off, but may now provide ordinary action authority only when the exact source projection, fresh held-prefix geometry, 4px lower kernel, active policy layer, interval, and version all match. Any unsupported source, delivery gap, expired layer, or empty predecessor still fails closed.

Removed lanes must not be re-enabled from archive without a new causal need and explicit STRATEGY.md decision.

Next Implementation Gate

TH08 work is active. Easy is the current diagnostic ladder; do not pay for a full route or re-enable non-authoritative global work yet:

  1. retain the source-driven Linux title/bootstrap state machine and its exact- child/no-duration contract; never dereference the dangling title global on the post-confirm release epoch;
  2. record a short ordinary Linux gameplay sequence and save its replay without writing gameplay or RNG state;
  3. define a compact semantic fingerprint with exact discrete fields and separately reported absolute/ULP float drift; stop on the first discrete divergence or threshold crossing;
  4. Done for one physical trajectory: replay a retained Windows-origin sample in both runtimes, then replay a normally saved Linux-origin file in a fresh Linux process and the original v1.00d Wine process. The aligned reverse trace is exact through natural game over;
  5. Done for a nontrivial Stage-1 prefix: connect a generic one-epoch Easy local planner and pass a 3,600-gameplay-epoch native smoke. The next gate is one full Easy route with no gameplay-duration cap.

Never unblock corridor action authority by provenance alone. The current active-spell request can omit future births, so scale, complete reached-source coverage, exact geometry, robust delay, and issue version must become one certificate.

Research Loop

Use this sequence:

first hit/root -> exact parent repeat -> offline/native branches -> first mismatch -> one general change -> same-root win -> native replay confirmation -> rotated focused physical trial -> full route after a major integrated gain

All-36 portfolios, long horizons, complete test suites, and full physical routes are milestone gates, not mandatory per-edit checks.

Commands

Import smoke

PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=scripts python3 - <<'PY'
import th08_live.controller
import th08_runtime.native_snapshot
import th08_runtime.native_snapshot_projection
import tools.th08_native_snapshot_causal_search
import analysis.th08_native_combat_branch_report
import th08_automation.practice_supervisor
import th08_automation.full_route_supervisor
print("import smoke: ok")
PY

Focused and complete Linux tests

PYTHONPATH=scripts python3 -m unittest discover -s tests \
  -p 'test_relevant_file.py'

PYTHONPATH=scripts python3 -m unittest discover -s tests -p 'test_*.py'

Windows UNC discovery

Do not use cmd.exe cd/pushd or ordinary UNC discovery. Use:

/mnt/c/Users/21992/AppData/Local/Microsoft/WindowsApps/python.exe -c \
  'import sys,unittest; root=r"\\wsl.localhost\ubuntu\home\pentester\coding\codex_ida\th08"; sys.path.insert(0,root+r"\scripts"); tests=root+r"\tests"; suite=unittest.TestLoader().discover(tests,pattern="test_*.py",top_level_dir=tests); result=unittest.TextTestRunner(verbosity=1).run(suite); raise SystemExit(0 if result.wasSuccessful() else 1)'

Change only pattern for a focused Windows run. Do not run Linux and Windows performance gates concurrently.

Native build

python3 scripts/tools/build_native_planner.py --target linux
python3 scripts/tools/build_native_planner.py --target windows
python3 scripts/tools/build_native_planner.py --target windows-x86

windows preserves the existing x86-64 target. Use windows-x86 for the 32-bit embeddable Python process that controls the PE32 TH08 executable under Wine; the runtime loader selects between them from Python's pointer width.

Use TOUHOU_DISABLE_NATIVE_PLANNER=1 only for an explicit rollback/ablation.

Source-stateful complete-stage fuzzer

PYTHONPATH=scripts .venv/bin/python scripts/build_th08_source_oracle.py

PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=scripts .venv/bin/python \
  scripts/analysis/th08_source_stage_fuzzer.py \
  --profile gate --seed 0xce0132 --count 1 \
  --planner-stride 30 --geometry-oracle-stride 60 \
  --output artifacts/benchmarks/th08_source_stage_fuzzer_gate_20260825.json

The command runs to the program's full frame count; it has no duration-based early kill. Use --replay for an exact stored program and --counterexample-dir --shrink-failures to retain and reduce failures. Exact coverage and nonclaims are in notes/review/TH08_SOURCE_STATEFUL_STAGE_FUZZER.md.

Physical trials

For this VPS's prefix-scoped Wine path, prepare and smoke-test via notes/operations/TH08_WINE_VPS.md before invoking the full-route supervisor.

Run only with explicit user authorization from Windows:

run_th08_practice_agent.bat --stage 3
run_th08_practice_agent.bat --stage 4a
run_th08_practice_agent.bat --stage 5
run_th08_practice_agent.bat --stage 6b
run_th08_full_route_agent.bat

The BAT wrappers add --armed. F8 starts, F9 stops, and F10 exits. Before launch verify TH08 identity, foreground, route/difficulty, gameplay state, no-life-decrement patch, and no-Bomb configuration. Monitor trace growth and the exact interop process; always release keys and clean up.

Common Traps

  • Different-RNG hit totals are not controlled A/B evidence.
  • A later hit is not independent after the first hit changed resources.
  • A replay future cannot be reused after a different action.
  • Native/Python parity is not model completeness.
  • Static shot width, damage, item availability, or Power gain is not survival improvement.
  • More tests, reports, schemas, or audited addresses are not solver progress unless they unblock a causal hit-reduction experiment.
  • Do not restore retired code because an archived note mentions it.