TH06 Hard Route Agent

This repository targets one physically validated Touhou 6 v1.02h Hard Reimu-A clear with no HIT, no Hard-authority loss, and no Bomb.

The active design is source-grounded route solving, not a universal online planner:

installed ECL + authoritative source + physical snapshots
    -> offline route/phase analysis and policy search
    -> short state-conditioned route intent
    -> common delivery-aware Hard certification
    -> one physical input transaction

Read START_HERE.md, then AGENTS.md and HANDOFF.md. The retired universal-planner checkpoint is preserved at annotated tag pre-phase-route-pivot-20260804.

Repository map

• scripts/th06/native.py captures one coherent native snapshot from the supported EXE.
• scripts/th06/model.py is the immutable runtime/source state model.
• scripts/th06/hazards/ contains route-neutral source physics and future world forecasting.
• scripts/th06/kernels/ and native/ contain parity-checked Hard/local hot paths.
• scripts/th06/solver.py is the small common authority and route dispatch.
• scripts/th06/routes/ contains route keys, source-phase identity, and the authored route packs. Hard/Reimu-A/Stage 4 is the first retained pilot; Stage 1 is now the active phase-by-phase route.
• scripts/th06/routes/state_machine.py contains isolated source-clock policy states; only the selected route phase can emit an intent.
• scripts/th06/barrage_lab/ provides source-derived/stateful offline workloads. It is an experiment bed, not a clear oracle.
• tests/corpus/counterexamples/ retains small understood shared-model and local-primitive counterexamples.
• notes/PHASE_ROUTE_PIVOT.md records the audit and architecture decision.
• notes/FUTURE_WORLD_AUDIT.md is the detailed historical source/model ledger.

Authority

The exact supported executable SHA-256 is:

9f76483c46256804792399296619c1274363c31cd8f1775fafb55106fb852245

The ignored authoritative source clone is reference/GensokyoClub-th06/ at cc475a0bc3fef38683b0f02224c87ddba0a021d9. Installed archives under reference/th06_dat/ are ignored and must not be committed.

Reverse-engineering claims come only from that source, existing IDA material, and controlled runtime/physical evidence. Do not use REA or LeanToken.

Checks

./build_th06_native.sh
./check_th06_baseline.sh

Inspect source timeline sections for route authoring:

PYTHONPATH=scripts python3 scripts/inspect_th06_route.py \
  reference/th06_dat/th06_ST.DAT --stage 1 --subroutine 14

Rebuild and hold out the Stage 1/sub14 compiled feedback tube from the retained physical entry snapshot:

PYTHONPATH=scripts python3 scripts/solve_th06_stage1_sub14.py \
  artifacts/th06_failure_stage1_f6358_sub14_compiler.json \
  --archive reference/th06_dat/th06_ST.DAT

Add --emit-python only when intentionally regenerating scripts/th06/routes/stage1_sub14_data.py. Training delivery seeds are kept separate from the default holdout seeds; neither is a runtime branch key.

Rebuild the Stage 1/sub12 residual policy from both stable ignored physical t61 workloads:

PYTHONPATH=scripts python3 scripts/solve_th06_stage1_sub12_residual.py \
  artifacts/th06_failure_stage1_f6110_sub12_compiler.json \
  artifacts/th06_failure_stage1_f5512_sub12_active_compiler.json \
  --archive reference/th06_dat/th06_ST.DAT \
  --entry-frame 6070 --entry-frame 5510

This compiler derives the feedback membership radius from the disjoint holdout. The compact policy is conditioned on source clock, position, held action, and the fresh Hard-4 action mask already computed by the common runtime. Physical frame and delivery seed are provenance only. Outside the measured radius the route retains the previously promoted right-stream policy; it does not extrapolate a nearest sample across the screen.

Screen the complete Stage 1 first-spell cycle from its retained physical local-t120 root (the workload stops at the life/timer callback boundary, not after an arbitrary short frame window):

PYTHONPATH=scripts python3 scripts/solve_th06_stage1_sub22.py \
  artifacts/th06_failure_stage1_f6273_sub22_uncovered.json \
  --entry-frame 6273 --seed-start 16 --seeds 16 \
  --candidates pv10

Run --candidates production after changing the route to exercise the real common Hard pass, deadline-aware route adapter, and publication ranker over the same candidate-conditioned battle world. Neither mode is physical-clear evidence.

Screen the complete Stage 1 second-nonspell RNG call family from the retained physical sub18 root. Success means reaching stable sub23, not merely surviving a fixed window:

PYTHONPATH=scripts python3 scripts/solve_th06_stage1_second_nonspell.py \
  artifacts/th06_failure_stage1_f8233_sub18_uncovered.json \
  --entry-frame 8233 --seed-start 0 --seeds 48 \
  --candidates production

The route keys policies by source subroutine, exact source-event segment, and the saved CALL caller where the initial sub16-to-sub18 entry differs from later recurrent sub18 entries. Delivery seed and physical frame are workload provenance only.

Run stateful offline replay from a physical snapshot:

PYTHONPATH=scripts python3 scripts/replay_th06_stateful.py \
  artifacts/th06_failure_latest.json \
  --archive reference/th06_dat/th06_ST.DAT \
  --seeds 500 --frames 240 --birth-events 4

For a captured physical battle world, --metric policy-volume replays the current bounded recursive primitive and --metric constant-frontier replays the unchanged-action frontier recovered from the historical clear. These are phase-policy experiments only; neither can change Hard eligibility or count as a physical result. --target x,y optionally reproduces the route's soft target tie-break inside the metric-preferred Hard set; omitting it runs the same primitive target-free.

Generic source-shaped barrage stress remains useful for shared primitives:

PYTHONPATH=scripts python3 scripts/stress_th06_barrages.py \
  reference/th06_dat/th06_ST.DAT --seeds 1000 --planner

Physical play

Active Stage 1 route:

run_th06_practice.bat --practice-stage 1 --seconds 300

Retained Stage 4 pilot:

run_th06_practice.bat --practice-stage 4 --seconds 300

Full-route validation:

run_th06_baseline.bat --seconds 1200

--rng-seed, --continue-on-failure, and short time windows are diagnostic only. Windows launch from WSL must be non-PTY. Every run must release input, stop its exact PID, and leave no game/agent/high-CPU process behind.

Current coverage

The common source model includes current bullets, lasers, enemy bodies, timeline enemy insertion, bounded ECL children, future laser mutation, RNG, and captured combat/resource state. Coverage is not equivalent to complete causal battle simulation: candidate-conditioned aim, damage/kill/retirement, callbacks, and resulting RNG/item effects still require phase-driven work.

The first Stage 4 route pack currently covers audited pre-boss timeline sections with route-selected, source-clock policy states and bounded local horizons. Boss ECL phases are deliberately reported as phase-unavailable until authored. This fail-visible boundary is intentional; no anonymous universal fallback hides missing route work.

Active development has returned to Stage 1 and will expose coverage one source phase at a time. Until a Stage 1 phase is authored and physically validated, its missing boundary remains visible rather than borrowing Stage 4 behavior or the retired universal strategy layer.