STRATEGY.md

Touhou Solver Strategy Ledger

Last updated: 2026-08-25.

This file records current action authority, not chronology. Evidence belongs in compact reports, run dossiers, and counterexamples.

Status

• Live: may affect physical input inside its declared boundary.
• Shadow: may observe/record but cannot change input.
• Offline: replay, native-root, model, or analysis only.
• Proposed: worth testing but has no action authority.
• Rejected/retired: must not be silently re-enabled.

Promotion requires one immutable causal hypothesis, same-root evidence, deadline-safe integration, a fail-safe fallback, rotated physical validation, and an explicit edit here.

The active physical objective is Lunatic Sakuya/Remilia Route-2 Final-B NMNB. Extra is deferred, and fixed-root wind-tunnel results cannot promote a change without a named original-engine physical mechanism.

Before spending another full Lunatic route, complete one Easy Route-2 Sakuya/Remilia NMNB diagnostic. Easy is an infrastructure acceptance rung, not a replacement target: every hit is analyzed against sensing, native state, geometry/future coverage, planning, issue latency, and delivered input. Search cost is not an adequate explanation for an Easy hit without profiler evidence.

Live

L1 — Robust no-Bomb survival

The promoted policy senses native state, decodes the complete active hazard pool, computes robust Boolean viability over declared timing/delay support, uses the baseline local beam with the promoted pre-loss continuation preference, applies a fresh issue-time collision certificate, and publishes through the exact-version action transaction.

Authority:

• survival is hard;
• no Bomb bit;
• signed-clearance geometry;
• active/held/pending input semantics;
• latest-version lookup and fail-safe fallback;
• Python/NumPy rollback for native implementation failures.

Limit:

• incomplete future hazards and resource/combat strategy still cause global kernel exhaustion;
• exact active/held/pending pipeline roots now reach planning and fresh issue when the observation invariant is available, but future hostile coverage remains incomplete;
• stage/spell identities do not define planner mechanics.

L2 — Final-B global-time-scale schedule

The revalidated native player-laser/time-scale recurrence may affect Route-2 Final-B only under the pinned static ECL identity and exact schedule authority. Content mismatch fails closed.

The constant-current-root continuation is diagnostic and unknown-direction. It may publish a queryable shadow global policy, but its target and action labels are stripped before planning/issue; it is not general live safety authority.

L2a — Stage 1–5 no-scale-writer schedule authority

Live inside its exact identity boundary. Restore spell global action authority without restoring the rejected constant-current-root assumption. The accepted boundary requires exact runtime/static ECL identity, the configured Route-2/difficulty/stage context, a coherent unit-scale root, a complete active installed-callback inventory, and a whole-file audit showing every 0x88/0x89 callback index is literal and none invokes or installs scale callbacks 18/28/29. Any mismatch fails closed.

Shipped native dataflow establishes callbacks 18/28/29 as the gameplay scale writers; the remaining writes are game/stage initialization resets. Decoded Stages 1–5 satisfy the static no-writer condition. Stage 6 and Final contain real callback-18 writes and are explicitly excluded.

Hard Stage-4A 20260801_184903 passed the named mechanism gate: 3,599 decisions were constrained after the configuration-invalid 183730 run had zero, all five reached spell phases received constraints, and recovery/ distant-recovery selections returned. It completed with 12 hits and zero Bombs. This promotes only finite unit-scale schedule provenance; it does not promote the fixed-grid corridor outcome.

The same physical trace exposed 57 coarse terminal-threat relaxations after an exact corridor authority had been named. Exact corridor and ordinary exact authorities are now non-relaxable. The follow-up 191508 recorded zero such relaxations.

L3 — Default-off ordinary prepublication predecessor

--ordinary-preexhaustion-authority remains default-off. Its former scalar boundary-reserve implementation is rejected and disconnected. The replacement uses player phase rather than the retained deathbomb-window value: phases 1/2 fail closed, while native movement phases 0/3 remain eligible.

For an ordinary nonspell at unit scale, it uses the exact observed active/held/pending pipeline root. The published ordinary policy is a 4px Boolean lower kernel whose required clearance includes the 2.828px cell radius. Every selected complete mask is certified across every pickup branch under a held/no-further-write continuation into the next queryable active policy layer. A command still pending at that layer must itself belong to the layer's safe action set.

Hard coverage must begin at the current observation and span the held prefix plus the remaining active-policy horizon under one exact hazard version. Missing prefix geometry, a non-4px policy, an expired/unqueryable layer, interval or version mismatch, and unseen birth/event coverage all produce no action authority. Coverage is not metadata-only: the retained projection's annular-sector emissions and hostile-body AABBs are consumed in the fresh prefix and in the 80-frame future kernel.

Observed-body early kill may rank only the resulting nonempty hard set plus the fresh issue-safe set. There is no objective fallback when the predecessor is empty.

Observed physical falsifier: run 20260731_152921 enabled the flag on 12,029 decisions but produced zero eligible/applicable/effective decisions. All 7,202 nonspell decisions failed the state gate. Native 0x0044AB40/0x0044C390/0x0044C650 shows player+0xE2A68 is the retained deathbomb-window limit installed on a hit, not a zero-when-alive active counter.

That wiring bug is not the only failure. Re-evaluating the physical roots with only the bad gate removed still permits down_left at global exhaustion frame 835. An uncontrollable no-pickup prefix makes all 17 actions share worst reserve zero at frame 850; playfield clamping repeats the all-action tie at the saturated corner at frame 910. The scalar code remains only as a deterministic counterexample and has no promotion path.

Future-source closure passes, but the former signed predecessor does not. At the real live H=80, its 16px terminal set is empty at all five retained roots; the earlier eight-action f817 result came from H=32.

The factorized lower replacement passes its deterministic semantic gate. Native collision-control projection v14 captures the ordinary manager singleton, active enemy main/auxiliary VMs, callback gates, emission descriptors, motion, phase, and timeline runtime. Reachable sources are executed fail-closed for 268 frames and lowered to bounded bullet sectors and hostile-body AABBs. Unsupported source semantics remain UNKNOWN.

Observation-aligned immutable replay roots 816/831/834/848/908 rebuild the retained f817/833/835/850/910 chain. Native exact recurrence and the live adapter agree that f817 permits only left_fast/down_left_fast, excluding active up_fast and issued up_left_fast. The later four roots are empty and fail closed. Windows measured f817 kernel/certificate delivery at 1412/21 ms; the five roots have no unresolved action. This is deterministic native-replay evidence.

Stage-5 physical runs through 20260801_115838 establish the current boundary. Offline/native hazard parity is insufficient because the live terminal query, sensor, and issue pipeline consume the action lease. The new fallback builds an H80 table indexed by final observed issue age. It models active input during computation, old pending evolution, held no-write, replacement by a different complete mask, pickup delay 0..6, action- conditioned future aim, and a fresh hostile-body slab before the final age read. Missing rows and nonpositive signed clearance remain fail closed.

Packed binary32 paths are bitwise equal to the scalar native-order oracle. Independent action predecessors are evaluated incrementally in the order of an already-observed local proposal; that proposal has no action authority and is reused only to avoid repeating stale-root computation. Fresh issue still recertifies all local actions, and the exact final-age row alone may override input. In 115838, 11/21 delayed scans were physically effective; median issue age was 25 frames versus 43 before proposal reuse. The run completed 18 hits, 10 nonspell, zero Bombs, accepted replay, and cleanup. Different-RNG hit counts remain observational.

The default-off successor now retains each effective delayed action as a versioned terminal-continuation lease. On every compatible fallback root it first computes the held no-write predecessor; if that is unsafe, bounded recovery directions may be evaluated independently, and a direction change requires a safe final-age row. The old lease bridges this computation only while exact binary32 position, active/held/pending support, epoch, stage, phase, unit scale, remaining horizon, fresh geometry, and its immutable content-addressed projection version remain compatible. Stage-route and spell-active are reread at final issue instead of inherited from capture. Any mismatch revokes it. Held no-write issue-age rows share one physical path, so renewal does not repeat 74 equivalent paths. The retained Stage-5 f1456/f1469 action is compatible at f1470; the former unconstrained f1479 switch is rejected. Empty actual rows at f817/833/835/850/910 remain empty. This is deterministic evidence, not physical promotion. L3 remains default-off until a rotated Stage-3 or 4A run exercises renewal/revocation. Local ranking remains computation order/objective only, never survival authority.

Stage-4A 20260801_130742 exercised lease creation/renewal but not physical consumption: six leases were created, 34 renewed, and zero were effective at issue. The first retained mismatch showed that native current-input publication at priority 17 is one frame ahead of priority-9 player movement. L3 lease v2 therefore separates published and motion histories and versions a TH08 publication-to-motion lag of one. It also replaces revocation on any fresh-body change with exact set containment against the active/future-body AABBs already consumed by the witness. An uncontained envelope still fails closed. Linux and Windows deterministic gates pass; this remains default-off pending a rotated Stage-3 physical gate.

Rotated Stage-3 run 20260801_134853 completed with one nonspell hit, zero Bombs, Power 128 throughout, an accepted replay, and cleanup. This is a useful fallback-path observation but not lease promotion: zero leases were created, renewed, selected, or effective, and all 4,725 ordinary-eligible roots lacked a future policy. Of 1,214 future-source attempts, 701 were rejected solely by a nonzero Stage-3 timeline fraction and 465 by a native clock-bracket crossing. The fraction blocker was over-conservative under already-required exact unit scale: native timer 0x00447421 preserves the fraction and increments the integer component, while timeline dispatch compares that integer. Source semantics v12 therefore accepts any finite fraction at unit scale and projects the same integer clock. Nonfinite timer state or nonunit scale remains UNKNOWN. A fresh Stage-3 run must demonstrate nonzero complete publications before its hit result can validate the lease.

The v12 repeat 20260801_141250 physically validates that semantic correction but not the route outcome. Complete source roots rose to 692/1,261 and drove 595 completed publications plus 265 effective exact ordinary issues. Four effective delayed rows created leases; all four were revoked by fresh-body containment and effective no-write lease consumption remained zero. The run had four hits versus the prior different-RNG one-hit result. At canonical hit f3181, the last exact issue was f1406 and every decision in f2941..3181 lacked a future policy. Thus the new global path was active earlier but did not provide persistent authority into the hit window. L3 remains default-off; do not attribute either aggregate to the fraction correction.

Parent-checkpoint cross-workload trials confirm that the Stage-3 one-hit run was not a general lease improvement. At 5fc6d92, Stage-4A 20260801_142851 completed 18 hits/10 nonspell versus the direct 130742 baseline's 21/12, but first hit was 833 versus 819. Stage-5 20260801_143952 completed exactly 18/10 versus direct baseline 115838, with first hit 1888 versus 1728. These different-RNG totals are observational. The general mechanism signal is stronger: Stage 4A created nine leases and Stage 5 created seven, yet neither workload consumed one effectively. Fresh-geometry revocation dominated both, matching all four Stage-3 v12 revocations. L3 remains default-off while active hostile native motion and fresh recertification are made causally consistent.

Source semantics v13 and lease v3 implement that correction without enlarging the viable set. Every contact-enabled active source body is advanced by the same captured ECL/native motion state as its emissions and retained as an exact root-to-horizon AABB trajectory. A fresh capture conditions only the observed body AABB at the capture frame; it does not restart a second linear future and compare two incompatible approximations. Issue recertification uses the stable native prefix snapshot at its own frame. Lease construction rejects older source semantics, while unstable snapshots, an uncontained observation, incomplete topology, or missing future coverage still fail closed. Focused Linux and Windows gates pass. Physical Stage-3 rotation remains required.

The Stage-3 rotation 20260801_152531 falsified lease v3 as a live authority, not source v13 as a whole. It completed six hits/four nonspell with no Bomb; different RNG prevents comparing that total causally with the v12 four-hit run. Mechanistically, 653/1,190 source captures completed, 267 prepublication and 13 delayed exact issues were effective, and leases were created ten times and renewed three. Two leases reached authority selection, but neither was effective at issue. Of 13 captured leases, six were revoked by bodies whose native contact gate was disabled, three by a decorated action/SHOT-only mask change despite identical direction/focus, and one by a redundant fresh-local veto after the exact old proof and observed body seam already passed. These are structural false revocations; fix them without widening direction/focus or future-source coverage, then repeat Stage 3 before rotating Stage 4A.

Lease v4 implements that narrow correction. Hard ordinary certificates and global submissions use only currently contact-enabled enemy bodies; the current exact source projection remains responsible for every future contact-enable transition. A lease consumes the held direction/focus mask and may survive only a SHOT-only auto-confirm write; Bomb, focus, or direction changes are still unequal. When a renewal table has no safe final issue-age row, the compatible older lease is selected before falling back to unconstrained local control. Stable contact-body seam containment and exact source coverage remain hard; the fresh local recertificate is diagnostic because it rereads no bullets and cannot supersede the old causal proof. Linux and focused Windows gates pass.

Stage-3 20260801_155718 physically validates this narrow mechanism: one lease create plus one renewal yielded 29 effective held/no-write issues, all with the exact certified down_left_fast movement and no physical input transition; the only revocation was terminal expiry. It does not validate the global outcome. The run completed seven hits/five nonspell, and every ordinary hit's preceding 240 frames had zero exact authority. Source prefixes were commonly truncated below H268 by reached main ECL 0x05/0x06 or captured movement state 2. Exact source closure, not another lease or scalar-reserve patch, is now the next authority gate.

Source semantics v14 implements that closure from shipped-code evidence. Main 0x05 decrements its captured integer lvalue and branches on the post-decrement operand; 0x06 writes its captured integer lvalue. Native state 2 is continued from stored displacement +0x2DC4, start +0x2DD0, easing bits, current/fractional timer +0x2DE4/+0x2DE0, and duration +0x2DE8. This also corrects the earlier state-3 timer read from previous +0x2DDC. No velocity inference or old-root uplift is allowed. Focused Linux and 147 Windows tests pass; Stage-3 full-horizon delivery is the next physical falsifier.

Stage-3 20260801_163434 passes that narrow semantic falsifier but exposes the next delivery boundary. Source-v14 completed 801/1,275 roots and 58 full H268 roots, up from 645/1,235 and 47 under v13; the targeted loop/set and timed-motion failures disappeared. However, all three ordinary hit windows still had zero exact authority. The first full root before f1253 was f1223, submitted at f1230, and its 2352.884 ms solve completed only at f1371. Thus v14 is retained as exact model progress, while the 4-versus-7 different-RNG hit delta has no promotion value. The next authority gate is earlier exact closure of newly reached 0x19, 0x6F, bottom-level 0x35, and resulting transform semantics; hidden operands and unavailable stack state remain UNKNOWN.

Source semantics v15 implements only the newly proved shipped cases. Main 0x19/0x1A reuse the set-valued captured-local arithmetic. Direct-fire type and color independently resolve their low/high dynamic i16 operands. 0x25 normalizes its float lvalue with a full [-pi,pi] union at a wrap seam, while depth-zero auxiliary 0x35 terminates; saved-frame returns remain UNKNOWN. 0x6F writes the exact selected 24-byte record and rejects an out-of-range index or set-valued transform field. Active vector/angular acceleration, stop/re-aim, and reflection are consumed as conservative full-disc paths; template replacement expands geometry to every reached target template. The actual Stage-3 time-65 main wave now deterministically closes through six fires. This earns a physical source-lead gate, not live promotion.

Stage-3 20260801_171249 passes that source-lead gate but falsifies the current recursive-delivery contract. Source v15 produced 83 full-H268 roots and removed its named failure reasons. The canonical f594 window contained 13 full roots and 43/50 effective exact prepublication issues, unlike every ordinary hit window under v14. At f565, however, down_left was proved only to terminal f571 and did not create a continuation lease. A new f566 root then spent 394.522 ms synchronously evaluating a constant-H80 delayed table; the game advanced 28 frames while the held input remained unchanged. Retained bullet slot 1 was already present at f566 and intersects that exact held path at f587. The automatic sensor_gap label observes only the later f594 row and is rejected by the retained causal reconstruction.

The next live change must treat computation itself as a held-input action. A long delayed scan may start only under an exact constant-hold guard or a compatible lease that covers its issue-age support. When no such guard exists, issue the bounded local/exact-prepublication result instead of blocking. Exact terminal probes should spend their bounded action budget on the held action and the current kernel's highest-repair viable alternatives before arbitrary compass defaults. This changes candidate computation only; the selected action still requires the complete prefix and terminal predecessor, and unknown coverage remains fail closed.

That correction is now implemented. The exact held-action H80 certificate is active as a computation guard rather than a compatibility trace. Without a compatible lease, the synchronous delayed table runs only when that certificate covers the complete effective issue-age and pickup support. If the held path is unsafe or its future projection is unavailable, the scan is skipped and no global authority is manufactured. The retained f566 slot-1 root deterministically fails this guard. Terminal candidate ordering now uses current-kernel repair volumes after the held action; the retained order is down_left/up_left/down instead of down_left/left_fast/right_fast. Focused Linux and Windows tests pass 33/33. Rotate the physical gate to Stage 4A and require the same guard/authority mechanism; aggregate hits alone are not acceptance.

The authorized Stage-4A physical run 20260731_220830 did not exercise that authority: all 1,754 live source projections were incomplete, chiefly because the former sparse capture crossed the manager frame 1,726 times. Consequently all 10,016 L3 records failed closed and zero selected actions were constrained. Its 23 hits and first hit 939 are different-RNG observations; they falsify the capture/delivery implementation, not the retained f817 action set or the observed-body early-kill objective.

Source semantics v3 now captures the contiguous manager+480-slot image in one read and lowers the reached dynamic-count, auxiliary subtraction, indexed-enemy timeline field, and FRScreen gate cases. The exact native kernel tests terminal membership first and publishes an all-action kernel directly when the complete clearance slab proves every clamped transition safe. Boolean hard membership omits soft scans; directional recovery is still computed when the predecessor is empty. Windows no-hazard viability is about 282 ms cold and 9 ms warm instead of the physical 2,996 ms bootstrap solve; the retained dense-root worst case is still about 2,104 ms and remains fail-closed on deadline expiry.

The correctly configured physical follow-up 20260731_231944 still produced only 1/1,859 complete source roots: 1,836 captures crossed manager frames and zero decisions received hard authority. 20260731_230657 omitted the runtime ECL flags and is configuration-invalid for this question. The next checkpoint retains one worker-local 10.32 MiB RPM destination and zero-copy pool views, removing about 10.6 ms of observer allocation/copy work from the bracket while preserving fail-closed semantics. L3 stays default-off until a fresh physical run shows nonzero effective exact authority.

That activation gate is now passed, but promotion is not. Stage-4A 20260731_233856 produced 119/1,914 complete roots and 196/193 applicable/effective ordinary constraints, proving the causal authority can control physical input. Before first hit 845, however, its 98 applicable roots were all trivial 17-action safe sets and 295 roots were unavailable; no nonempty effective authority occurred within 80 frames of any hit. Source semantics v4/projection v13 now lowers native motion-angle variable 10069, the complete auxiliary timer root, and generic opcode-0x02 delay scheduling. Armed phase-successor coverage remains the next fail-closed blocker.

Stage-4A 20260801_002006 physically falsified the remaining publication path: despite complete source roots 1137..1200, an earlier UNKNOWN-source kernel occupied the serial worker/pending slot and exact authority was absent from all 383 decisions in the 80-frame nonspell hit windows. L3 now refuses to solve an incomplete ordinary source and admits a completed exact pending future policy as the causal predecessor's terminal kernel before activation. This preserves fail-closed coverage while removing work and publication that could never carry hard authority. Physical revalidation is pending; phase, transform, opcode, callback, and movement closure is still incomplete.

Stage-4A 20260801_004533 validates that delivery correction but not L3 promotion. Completed pending policies supplied 110 effective nonspell decisions, including 43 before first hit 1837. Yet exact authority remained absent from every 80-frame nonspell hit window. Complete sources carried up to 750 future sectors; representative long solves spent 6.1..23.2 seconds in clearance construction and only 0.23..0.81 seconds in Boolean recurrence.

The native annular-sector kernel now groups only exactly equal origin/radial/inflation samples, forms their exact circular angular union, and evaluates one distance per group/lattice point. Disjoint frame slabs run over at most 16 workers. This changes neither the set-valued hazard union nor the signed-clearance threshold. Synthetic 750-trajectory pressure and randomized wrap/overlap/disjoint cases match the independent Python oracle; Linux, Windows, and the retained five-root semantic chain pass.

Stage-4A 20260801_011902 physically validates that geometry correction: maximum exact clearance time fell from 23.2 seconds to 0.993 seconds. It does not promote L3. Authority was effective on 399/7,162 nonspell decisions but on 0/380 decisions in 80-frame nonspell hit windows. The canonical frame-796 root spent 478 ms in clearance and 918 ms in Boolean viability; publication missed the last nonnegative corridor gate at frame 714, and the predecessor was empty by frame 716. The 1,195 armed phase-transition source failures are an aggregate later blocker, not the causal explanation for the first hit.

The Boolean kernel now proves common all-action interior states through a set-valued box containing every held/pickup branch at every physical step. The shortcut applies only when every covered signed-clearance cell and every terminal action is safe; all other states execute the unchanged exact branch recurrence. The retained Windows chain preserves exactly left_fast/down_left_fast at f817 and empty sets thereafter, with no unresolved action, while maximum solve time falls from 2,071 to 775 ms. Physical revalidation is pending. Missing source coverage still fails closed.

Stage-4A 20260801_015631 physically validates the recurrence optimization without closing L3. Nonspell exact solve maximum fell from 3,050 to 1,338 ms; viability maximum/p95 fell from 3,026/1,664 to 1,038/926 ms. Exact authority was applicable/effective on 494/479 of 6,941 nonspell decisions, and 102 of 278 pre-first-hit applicable decisions were directional. It remained absent from all 238 decisions in the 80-frame nonspell hit windows. The complete first-hit-window failure is now source coverage: roots 1915..2007 chiefly reached a nonzero bullet transform program, with opcode-0x19 failures mixed in.

Shipped bullet_apply_next_transform confirms a record executes only when its kind intersects the bullet's original flags. Source semantics v5 retains that gating. Active player-relative stop/re-aim (0x80) is represented by a full-direction path-length disc bounded by the maximum emitted/resume speed; it does not condition on an unobserved future player. Timed barrier, offscreen-cull suppression, and sound records are movement-neutral. Other active transform kinds remain UNKNOWN. Auxiliary float addition (0x19) is advanced with interval arithmetic. Retained Windows action sets are unchanged; physical pressure-window revalidation remains required.

Stage-4A 20260801_022228 confirms transform/0x19 closure: neither failure class remains in 1,952 live projections. Three decisions in the frame-3568 nonspell hit window obtained directional exact sets of 5/8/6 actions, the first nonzero pressure-window applicability. They were not effective at issue. Post-capture advance was 15/16/12 frames against pickup-support high 6, so the deadline guard preserved held up_right; at frame 3501 that held action was outside the five-action exact set.

This is now a causal delivery/resource-isolation failure, not scalar reserve, transform coverage, or local ranking. Ordinary background solves run at below-normal owner-thread priority and use eight native workers rather than sixteen, reserving CPU for the game, sensor, and issue controller. Windows retained solve maximum changes 804→1,095 ms while the exact five-root sets are unchanged. Reached auxiliary loop-jump 0x05 and integer-LE jump 0x2E are also lowered from exact captured integer state. Physical revalidation is required; unhandled source flow still fails closed.

Stage-4A 20260801_024419 falsifies the QoS change as sufficient. The eight-worker/background-low-priority path was active and ordinary solve p95 improved 230→218 ms, but one solve took 5,290 ms; the two directional exact sets in nonspell hit windows both arrived 20 frames after capture, leaving effective directional authority at zero. Aggregate 22 hits are observational.

The run also isolates phase coverage: 0x05/0x2E failures are gone, while ordinary slot-0 armed transition accounts for 1,243/1,714 incomplete source roots. IDA confirms timeout uses integer phase timer >= deadline, and health uses signed HP < threshold; both start the registered successor. Source v6 captures the four health successors plus the phase timer and proves only a timeout with elapsed + H < deadline unreachable. Armed health and reachable timeouts continue to fail closed. Retained exact sets are unchanged; physical revalidation precedes Stage 5.

The 2026-08-25 offline successor closes the next generic callback integration boundary without promoting L3. Retained child/timeline births and ordered callbacks 12/14 now reach a source-derived current-pool composer. A versioned certificate binds projection root, point-valued bullet snapshot, policy epoch, complete horizon, and the actual composed pool; global geometry v4 shares the 21-type spawn lifecycle/callback schedule and carries a current-transform completeness bit. CorridorSolution carries that join through solve and an independent action-authority check. The controller attempts it only under exact unit scale and zero bullet-capture span, and an expensive pool composition starts only when a submission is due and its worker slot is free; an incomplete join consumes no worker slot. A worker-root certificate cannot authorize a later local prefix over newly sensed bullets.

The first intended physical same-root experiment found two sharper boundaries. All four Stage-5 capsules are v1 producer roots without current bullets/lasers, so they cannot be completed by pairing another pool. New explicit retention uses a clock-joined v2 compact current-hazard root and passes a 319-bullet, two-laser stateful replay. Separately, the generic fuzzer proves the historical active-transform growth heuristic unsound: 1,069/1,969 samples escape, worst 97.060802px versus 31px. Such roots are now shadow-only. Coherent diagnostic roots now retain the exact 18-record program and every meaningful active handler block without changing packed local sensing. Stateful schema v4 also separates collision geometry from the ANM visual geometry used by reflection and culling for all 21 real templates. Therefore this remains offline/default-off infrastructure inside L3. Source-order Python/C execution and the batch differential now pass; template/derived births and global consumption remain open. The next physical falsifier is a complete Easy Route-2 diagnostic, followed by a fresh v2 Stage-5 same-root global-policy change after those remaining semantics close. No deadline, Wine action, or hit reduction follows from the offline kernel. Complete Linux discovery passes 1,544 tests with five conditional skips, and the paper builds to nine pages.

L4 — Continuous-position coarse-cell lower bound

The live 16px global corridor no longer treats a lattice center as the whole state. Its required clearance consumes the cell half-diagonal sqrt(2)*8 = 11.3137px; exact corridor authority may not relax to a local coarse fallback. This is a semantic hardening, not an outcome promotion.

Hard Stage-4A 20260801_191508 physically falsified the uniform coarse lower kernel as a useful global solver. Relative to different-RNG 184903, empty queried sets rose 1,723→2,927, constrained decisions fell 3,599→3,001, boundary-associated hits rose 5→12, and the run completed 18 hits with zero Bombs. Query availability and median solve latency improved, so missing publication alone does not explain the regression. Retain the sound lower bound, but require adaptive/local exact refinement before any outcome claim.

Offline Native Authority

O1 — Rolling native snapshot wind tunnel

Original TH08 may be held at a canonical root, branched over complete no-Bomb actions, advanced through the original calculation chain, verified, and restored in one warm session.

Accepted evidence at Stage-5 root 2129:

• parent replay exact;
• all 36 root masks;
• 324 causal branches;
• H=32 witness 0x94 -> 0x44 -> 0x10 -> 0xA4;
• native survival through frame 2161;
• 32/32 natural-frame-pump state agreement.

This is exact only for the declared replay/root/horizon. It has no live delivery or whole-stage authority. A persistent warm service remains proposed until mapping-epoch poison recovery, single-writer ownership, cancellation, TTL, and rebootstrap are implemented without weakening checks.

O2 — Fail-closed model/native differential

An explicit model may advance only supported player, hostile, enemy, combat/resource, timing, and RNG events. It must stop at the first unsupported mutation or bit-relevant mismatch. A recorded future is never reused after action divergence.

The current unresolved canonical mismatch is the pre-enemy/pre-aux producer before the known auxiliary fires at root 2129. Retrospective RNG alignment is diagnostic, not causal closure.

The retained-current transform prefix now has independent scalar and C executors. They cover queue/immediate semantics, all source-order motion handlers, barrier/wrap, movement, visual culling, and retirement; template replacement and derived finite-pool births remain typed UNKNOWN. At 1,536 states the persistent native kernel is 0.02446 ms median, versus 22.233 ms for Python-object encode, one frame, and decode. Global work may exploit the kernel through persistent flat state; the local issue loop may not claim the object wrapper meets deadline.

O3 — Route-2 combat/resource reconstruction

The current WS-H model can select generation-safe roots and compare completed native transactions for normal shots, enemy HP/defeat, phase transitions, items, Power, and resources. The general model remains offline. One narrow, default-off live experiment may align with an observed ordinary non-boss enemy only after an exact ordinary viable action set and a fresh issue-safe set exist under hard no-Bomb. The rejected scalar reserve and unconstrained interior are not eligibility sources. This is objective ranking, not independent safety authority. See notes/CURRENT_COMBAT_RESOURCE_MODEL.md.

Proposed Research

P1 — Survival-filtered kill-before-saturation

For ordinary enemies, compare earlier native kills and prevented later hostile births only among actions already feasible under the survival constraint. Test across more than one root/stage before promotion.

Preserve this objective, but exact ordinary-stage global action authority is now the higher-priority dependency.

Current evidence is mixed. A same-root Stage-5 branch observed one 20-HP enemy defeated and three additional hostile bullets suppressed every nine frames, while the current offline global viable/safe-action masks remained unchanged at three later frames. The first default-off physical gate applied 27 fresh-safe preferences, had first hit 6981 and 13 total hits versus a different-RNG 2124/12 baseline, and received no global policy guidance. This physically exercises delivery but neither promotes nor falsifies the general hypothesis. The zero-publication cause was the diagnostic scale horizon (120 available versus 161–162 required), not kernel exhaustion or player projection. The rotated Stage-4A gate then delivered 1,915 unique policies and 12,156 available queries. Of 39 actual fresh-safe early-kill choices, 26 occurred after the shadow kernel was already losing and only nine were also in a winning shadow-global action set. This preserves the strategy but rejects short local safety as its global survival filter.

The next Stage-4A physical gate enforced that rejection: 124 preferences were applied before the canonical first hit, each inside a winning queried shadow-global set and then a fresh issue-safe set. It completed with 11 hits, first hit 4148, and zero Bombs versus a different-RNG 16/1827 predecessor; that aggregate is encouraging but observational. The causal counterexample is sharper: the last winning query was frame 3679, while a 200-HP middle-wave enemy was body-visible near x=320 at frame 3864 but was not selected by the old HP gate until frame 3900 at 15 HP.

The default-off correction recognizes full-health ordinary non-boss enemies. The physical 20260731_142342 gate preserves the observed-body part but falsifies the forecast implementation: all 376 observations recycled the timeline-0 time-1 x=30 startup birth and only three affected input. A zero instruction pointer is currently aliased to timeline start without causal not-started/completed lifecycle identity. The forecast is now removed from the live preference path until that identity is exact.

P2 — Focus as a dynamic action factor

Do not hold Shift permanently. Include focus/unfocus in action generation:

• focus for precision and the corresponding shot pattern;
• unfocus for faster movement and alternate coverage when precision is not needed;
• allow causal refocus schedules.

Static width/damage tables are root selectors, not a policy. Same-root native survival and combat consequences decide.

P3 — Survival-feasible early Power collection

At clean early-stage prefixes, allow collection paths only inside the viable set. Promote only when a native transaction crosses a Power threshold and a later combat/survival join improves. Later dense phases remain survival-first. Post-death Power recovery is outside NMNB.

P4 — Hierarchical robust solver

The intended general architecture is:

1. hard robust viability/collision filter;
2. action-factor expansion over direction, hold duration, and Focus;
3. inside the viable set, short-horizon objectives for interior clearance, useful position, nonspell damage/exposure, and Power;
4. exact or conservatively bounded refinement at canonical losing roots;
5. native-root validation and live deadline-safe delivery.

Prefer proof-backed feasibility, dominance, canonicalization, and admissible bounds. MCTS, learned value functions, Monte Carlo, beam widening, and imitation may rank or propose candidates offline, but cannot prune hard-safety branches or gain live authority without an independent exact verifier.

The first scalar pre-exhaustion implementation is physically rejected. Stage-4A run 20260731_152921 produced zero live constraints because of an incorrect native eligibility predicate, and its retained roots show that a predicate-only repair still permits the losing action and aliases every action under prefix/boundary saturation. Contact bullet slot 455 first entered the retained nearby set while an older hazard-snapshot policy certified down_left; the next snapshot reported an empty kernel. The compact trace cannot attribute the version flip to that slot alone. The factorized 4px lower predecessor and bounded ordinary future-source coverage now pass the retained semantic gate. The next question is physical delivery/effectiveness on Stage 4A, not another boundary score or local ranking change.

P5 — Representative root corpus

Expand the wind tunnel from one Stage-5 first hit to a small diverse corpus: Stage 3, 4A, 5, and Final-B; ordinary enemy and spell; early Power and dense survival; different hazard/event classes. Root diversity is more valuable than extending one root indefinitely.

Rejected Or Retired

• Supplemental local lane: retired. It duplicated decision paths and its Windows delivery/contention work did not establish live benefit.
• Candidate verifier and pipeline prewarm services: retired. Dormant publication complexity without current hit-reduction authority.
• G5 auxiliary/bullet-birth program: retired as an active workstream. Its useful causal lessons are folded into the wind-tunnel/model boundary.
• Priority-17 publication and stationary-witness delivery lanes: retired.
• Old focused Final-B trial/report tools: retired. The exact scale schedule remains in the main full-route path.
• Scalar minimum boundary reserve as ordinary global authority: rejected by physical run 20260731_152921. It never activated because of a stale native-field predicate and, after counterfactual gate repair, still degenerates to all-action ties before the canonical hit.
• Replay-prefix exhaustive branching as inner loop: rejected for speed; native replay remains a sparse independent oracle.
• Action-incompatible replay future reuse: rejected as noncausal.
• Static “unfocused is wider/better” ranking: rejected by the Route-2 SHT atlas.
• Active-only enemy disappearance as kill evidence: rejected; end reason requires lifecycle and damage ordering.
• Different-root aggregate hit deltas as promotion evidence: rejected.
• Stage/spell-specific planner patches before general mechanics: deferred unless a repeated physical counterexample proves necessity.

Promotion Sequence

For any proposed strategy:

1. name the global hit/solver failure it addresses;
2. define its causal root, observable state, actions, uncertainty, and metric;
3. reproduce the parent;
4. show an offline/native same-root win or model-mismatch closure;
5. repeat on another representative root;
6. integrate default-off and prove deadline/fallback behavior;
7. run a rotated focused physical stage;
8. repeat before live promotion;
9. after a major integrated gain, run a fresh full Lunatic route.

If a step fails, record the counterexample and return to the responsible layer. Do not hide the failure behind more tests, reports, or stage tuning.