This is an experiment order, not a framework roadmap. HANDOFF.md identifies
the exact current step.
- Decode the installed Stage 4 timeline directly from
th06_ST.DAT. - Verify stable route and source-phase identities in physical snapshots.
- Keep the common runtime limited to fresh Hard, delivery, dispatch, and publication.
- Remove universal effort-ladder/guidance behavior and its tests from the main path.
- Run default fail-close Practice Stage 4 and stop at the first phase failure.
Exit: the first physical failure names an exact route ID and source phase, and no old planner silently supplies strategy.
For the earliest unresolved phase:
- audit its timeline spawns and ECL subroutines;
- seed barrage_lab with the physical entry distribution, active bullets, enemies, RNG, Power, and future source events;
- compare a few small phase-appropriate algorithms/parameters;
- compile the smallest state-conditioned policy or native helper;
- Hard-filter it online and rerun physically.
Horizontal crossing waves are an early priority. Test them as sequences with entry state and future births, not isolated stripes.
Exit: default Practice play crosses the phase for varied ordinary RNG without HIT, authority loss, Bomb, or stale publication.
Add candidate-conditioned aim, shot damage, enemy kill/retirement, callbacks, items/Power, and RNG ordering along the exact Stage 4 source paths that change route decisions. Verify each new transition against source and adjacent physical frames. Do not build a complete simulator ahead of the next phase.
Exit: offline branches no longer borrow impossible hostile births or omit kill-dependent route/resource outcomes for the owning phase.
Identify phases by boss ECL subroutine index plus callback/spell state. Give each nonspell/spell its own algorithm and parameters. Search entry position, damage alignment, timeout/kill branch, and recovery state offline; keep Hard authority common.
Exit: a default fail-close Hard/Reimu-A Practice Stage 4 reaches its source-defined result path with zero HIT, authority loss, and Bomb.
Author Stage 1, then carry actual Power/lives/RNG/position into Stage 2 and so on. Practice isolates phase work; full-route runs validate resource and entry state. A stage policy must accept a distribution of valid entries rather than one prerecorded state.
Exit: every stage passes both focused Practice and the current full-route prefix.
Run from the game start with default RNG and fail-close control. Success requires no HIT, no Hard loss, no Bomb, and the source-defined result/replay path. Forced RNG, continue-on-failure, offline replay, and patched-life survival do not count.