[dayfine]

Depends on #219

Adds the next two FRLG AutoStory rows after 00.20. 00.21 goes from Cerulean's Pokemon Center through the burgled house (the only way south without Cut, so the Rocket in its yard is fought), then Route 5, the Underground Path and Route 6 to Vermilion's Center; on Route 6 it walks down column 15 through the grass on legs to walls, out of Keigo's and Elijah's lines, and fights only Jeff, who stands across the one way south. 00.22 shows the S.S. Ticket at the harbor, fights the rival below the Captain's door, gets HM01 from the Captain and walks off as the ship sails, back to the Center; Cut isn't taught yet. Along the way the map generator reads tileset paths from metatiles.h (the S.S. Anne's didn't follow the symbol), a "new_game" flag set hides the S.S. Anne rival, and story routes no longer plan Surf, which they did across Vermilion's harbor. Tested with new OverworldMaps and AutoStory Simulation tests and the existing AutoStory/Overworld suites, all green locally; nothing has run on the console yet.

WhiteDialogDetector now also requires the dialog's blue top frame, so the S.S. Anne's white hull no longer reads as a message and stops a walk (console B, 00.22); new WhiteDialogDetector image test, frames in dayfine/CommandLineTests#10.

The win-chance sim now plays the battle loop's in-battle Antidote and logs how its battles are lost (foe out, seeded/poisoned/asleep, accuracy down, out of Potions); the S.S. Anne rival's 40% is real for Bite/Water Pulse/Bubble/Withdraw, lost to the Ivysaur's Leech Seed (new AutoStoryBattleSim test).

The battle loop no longer takes a send-out flash for a move prompt's YES/NO (it needs the box's frame now; m68's NO then YES set the lead to Lv31, 88 max HP for 78), takes a lone "no PP left" as a Disable, and at high risk uses the win-chance sim's own Potion rule (poison, likely Leech Seed, the foe's HP bar), which also tops up before a knockout; S.S. Anne rival 31.4%/38.5% with an Antidote, about unchanged, most losses still to the seeded Ivysaur (images in dayfine/CommandLineTests#11).

Rows 00.23-00.25 continue to Lt. Surge: HM01 Cut to the Oddish and TM28 Dig over Withdraw, a new CUT_TREE step for Vermilion's gym tree, Dwayne, Baily and Tucker (the two who look around are met by a step onto their line's second tile and a wait), the trash cans searched by counting each can's message pages (a wrong second can resets, and the search goes on from there), a Route 6 GRIND until the sim beats Surge 50% (Dig takes it from 6% to 47% at Lv28), a save in front of Surge, and Surge; tested with new AutoStoryTrashCans, AutoStory Simulation (00.23-00.25 walks, every can-to-can walk past the trainers where they stand after their battles) and OverworldMaps tests, nothing on the console yet.

The summary's KNOWN MOVES reader now knows the Oddish line's moves (it read none of the Oddish's, so 00.23's Cut teach stopped as unread) and ends the moves at the first slot with no type badge, so a move it can't name keeps its row instead of cutting off the rest; new SummaryMovesReader image test, frame in dayfine/CommandLineTests#11.

The TEACH step now goes through the TM's own scene ("Machine set!", " learned !", pokefirered pokemon_special_anim.c) with A and never B, which cancels the teach; 00.23's Cut onto the Oddish stopped there and gave up. Free slot and forget path, TM and HM, are covered by AutoStoryTeachScreens sim cases and a new TeachScreenReader image test (frame in dayfine/CommandLineTests#11).

On the forget path the step now presses A on every look of the scene until the TM CASE returns: console B read the forget's "Poof!" box (which waits for A) as "learned" and stopped pressing for 5 minutes on TM28 Dig; leave() waits for a scene once, so a stuck one gives up instead of looping (AutoStoryTeachScreens sim cases).

An open leg over a map connection, held a tile longer for the game's stall there, can end a tile further when the game doesn't stall (00.24, Route 6 south into Vermilion: north 1 onto the Center's door stopped below it); a warp walked back that way is now held a tile further, stopping at the fade (OverworldSim test walks the exact plan with and without the stall). 00.24 also checks Lt. Surge's win chance in the Center first and skips the trip to Route 6 when it already passes (AutoStoryGrind and 00.24 sim tests).

The Lt. Surge loss (m73, sim 79%) came from the Raichu's Double Team at full HP and Dig's two turns, neither in the sim; it now scores Double Team as pret's AI_CV_EvasionUp and plays two-turn moves (charge turn, under ground), giving m73's lead 58% (82% at Lv31) (Vermilion Gym sim tests).

Story action table (core stage 3, from #213): the GAME if-chain in StoryExecutor::game() is now four table rows (close_trainer_card, leave_gift_summary, cut_tree, trash_cans) plus the GAME default for captures, PC and abilities; the executor has no per-kind virtuals left. The AutoStoryStepTable parity tests check each row against the if-chain it replaces.

Sim on the story action table (core stage 8): the full-run simulation and row_step_on_sim() no longer dispatch by kind themselves. StoryMapsSim (the whole story on the story maps, formerly the test's FakeGame) and OverworldStepSim (a row's steps on the OverworldSim; row_step_on_sim() now runs through it) bind make_story_action_table() and run through TableExecutor, stubbing the actions that need the console; stub_report() lists each row's stubbed steps, and a new AutoStory Simulation test checks the whole story stubs exactly the steps whose action needs the console. Checked at #258's tip: the full-run step trace (1395 lines incl. outcomes and saves) and all 17 StoryReplay reports are byte-identical, and the FRLG and Pokemon:: results are unchanged apart from the new passing test.

Interrupts between steps in the sim: StoryMapsSim binds the same table; StorySimKnobs::dialog_before pops up a dialog before a step, which stops a walk or a talk until it is cleared, and clear_interrupts = false keeps the old behaviour. Two new AutoStory Simulation tests: a dialog before every step (red before the binding), and dialogs before one step in five over 5 seeds, 0/5 runs through before and 5/5 after. Full-run trace and replay reports unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AKYzhrMnPK61XHWEch4a6E