[dayfine]

Depends on #258 (base frlg/checkpoint-replay; the diff shows only this change).

The table stopped at Route 10's Pokemon Center (00.27). 00.28 cuts the tree (10, 20), goes through Rock Tunnel without Flash, fights the nine trainers the way can't go round (Cooper by the walker on a leg to a wall, as Mt. Moon's Grunt) and saves on Route 10 by the south mouth. 00.29 goes down Route 10's east side past Carol, Clark, Trent and Herman, hops the ledge (11, 74) and walks to Lavender Town's nurse. Tunnel legs end on walls, ladders or people, and open legs are followed only by legs that a tile of error can't send astray (the 00.27 lesson). No row uses a Repel, because the table has no step for it. The map resource lock moves to PokemonSimulation 6c2f5c8 (dayfine/PokemonSimulation#80), which adds RockTunnel_1F/B1F, LavenderTown and its Center from the decomp. The only change there to an existing map is Route10's links to the new maps (no tile, elevation or object changes).

Testing: there is a new AutoStoryLavender sim suite. It runs each row whole, starts the rest of the row a tile short or long after every open leg, and adds a wild battle mid-leg in the tunnel. The FRLG AutoStory filter passes except the 12 known FileException image tests, StoryReplay passes 18/18, the step-table lint finds no unbound step, and code health has 0 failures. Nothing was checked on a console. The detectors are unmeasured under Rock Tunnel's darkness, and the plan for measuring them is in ~/FRLGLab/notes/2026-11-09-0028-dark-measure.md.

🤖 Generated with Claude Code

https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss

[dayfine]

FRLG movement mechanics are now one table, Overworld/PokemonFRLG_MovementMechanics.h. It holds the pace, the turn and step taps, the releases, and the tap rules measured on console B on 2026-11-09. The walker, step_once() and OverworldSim all read it.

• Sim: a tap from a standstill only turns a player who faces another way. It takes one step if they already face that way. A step tap that starts within 620 ms of the last step tap's start is ignored. Taps with no release between them count as one hold. The random 1-in-10 turn drop and its MC case are gone. dropped_turn_taps is kept only as fault injection for the turn_to() facing check.
• Walker spacing: step_once() tapped 120 ms and released 500 ms, so its taps started 620 ms apart. That sits in the unmeasured band between 566 ms (ignored) and 666 ms (worked). It now taps 66 ms and releases 700 ms, 766 ms apart. One-tile legs come at least 666 ms apart (run) or 733 ms apart (walk) because of the 500 ms after-leg watch, so they are unchanged. hold_to_warp() can start its long hold inside the lockout, but the hold has two tiles of slack.
• Tests: OverworldWalk "tap:" cases replay the console measurements, plus three step_once() calls in a row. That case was red with the sim's old 340 ms tap and is green now.
• 00.28 MC, seeds 1–30, game odds: before 30/30 passed, time mean 920.7 s / p90 1008.2 s, tiles walked again mean 5.8 / p90 52.0. After 30/30 passed, mean 938.0 s / p90 1026.8 s, tiles walked again 5.8 / 52.0, and 0 step taps ignored. The added time is the step tap spacing, now 766 ms where the sim used to tap every 340 ms.

[dayfine]

00.28 steps 14/15 (console B 13:35 failure): fixed in f6080f0 + 79bfa97.

• Rock Tunnel's dark overworld (black corners, lit middle; dark_cave_screen()) is no longer read as a battle on its way, before or after a step, or by the screen observer. The thresholds are not measured on console frames yet.
• An open leg that starts in a dark cave counts a stop as anywhere on the leg (stop_tiles() → whole_leg()). This replaces the timing, which placed the battle at the end of east 12 at 0–3 tiles in. plan_reanchor() already says "found" only when every candidate ends on the same tile, so the 5-tile miss came from the dishonest candidate set.
• below_martha_line is now west 16, north 2, west 2 to the corner (11, 35), east 5. After a battle, the corner finds the player from any of (11..16, 35).
• Sim: SimScript::dark_maps, walker rules dark_cave_not_battle and whole_leg_in_dark. New tests: "B1F dark, the 13:35 legs" (red without the rules on 13 of 53 battle steps, green with them) and "B1F dark, a wild battle on any step to Martha's line" (steps 1–40, on time and late). The game-odds MC now walks the tunnel dark.
• MC 00.28, seeds 1–30. Before (old code, sim without darkness): 30/30, time mean 938.0 s / p90 1026.8, tiles walked again mean 5.8 / p90 52. After (dark modelled): 30/30, mean 1023.4 s / p90 1177.1, tiles walked again mean 61.7 / p90 118.

Not done: the measured north tap. The corner clip covers row 35. Step 15 accepting a bump into Martha is also not done.

[dayfine]

db14a25: fix for 00.28 hiker_lenny (lab-B 2026-11-09 15:11). After a win in Rock Tunnel without Flash, the dark overworld read as the battle text box with no HP box, so the loop held B for an "evolution" 3 x 20 s and then gave up (no START menu for 60 s). Now a dark_cave_screen() frame reads as no battle screen and EvolutionHold never holds on it, so START finds the overworld within seconds.

Red first (2b8b70e): a decisions test, an 00.28 dark sim test, and a new DARK_CAVE fault sweep (Rock Tunnel dark, with no wild battles and with 10 seeds at the game's odds). The sim now plays the battle loop's end after every trainer battle, including the walker's. Tests: PokemonFRLG::AutoStory 363 passed / 13 failed, the same 13 as the baseline (11 AutoStoryBattle images, eng_chansey, Resources). Code health: 0 failures.

🤖 Generated with Claude Code

https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss

[dayfine]

5e00efa on #213: FRLG's look between steps (ConsoleScreenLook) now goes through Pokemon::SwitchHomeObserver, so a step that starts on the HOME menu fails as game_not_running before any input (the same as BDSP, from #122). Merged down to here (a22824f). Local headless runner: PokemonFRLG::AutoStory 363/13 (the known image cases and AutoStoryResources), Pokemon:: 210/0.

[dayfine]

413f1cb: 00.28 winston_line lost its 3rd tap after a wild battle on the 2nd (console B 15:50). Taps after a battle run from mid-line now wait until the overworld has shown nothing for 1 s (wait_out_battle_exit() over the shared ScreenLook/ScreenSettle), as does capture pacing. The sim now ignores taps for 0.5 s after a battle. Added the red regression test "00.28 B1F dark, winston_line: a wild battle on the 2nd tap" and a LOST_TAP fault sweep (a battle on each tile of each trainer line before its last). Not built or run locally: disk was below 32 GB. Code health and a syntax-only compile of the 7 changed .cpp files pass. Mac and ubuntu-clang CI dispatched.

[dayfine]

Built and tested 413f1cb locally (headless SerialProgramsTests, upstream/headless-tests merged in a scratch tree only). Red check: with sim_wait_out_battle_exit commented out, winston_line fails ("D-pad held west ... in the lockout after a battle"); restored, it passes. --filter PokemonFRLG::AutoStory: 399 passed / 13 failed, the 13 being the known local set (11 AutoStoryBattle images, WildEncounterReader eng_chansey.jpg, AutoStoryResources). Both LOST_TAP sweep cases (00.28, 00.29) pass. Code health: 0 failures. No code changes needed.