Skip to content

FRLG: emit [Leg] records, replay the 12.03 unmoved-hold case - #451

Draft
dayfine wants to merge 16 commits into
frlg/checkpoint-replayfrom
frlg/leg-records
Draft

dayfine wants to merge 16 commits into
frlg/checkpoint-replayfrom
frlg/leg-records

Conversation

@dayfine

@dayfine dayfine commented Nov 13, 2026 •

Copy link
Copy Markdown
Owner

Depends on #258, with #464 (the shared leg trace and replay) merged in; the diff still shows those files until #258's chain has them.

FRLG logged only prose for a walk, so a hold that moved nothing (12.03, Pallet Town: both east holds from (6, 10) moved 0 tiles, logged "ended at (10, 10) (walked)") could not be told from one that worked. The shared Pokemon_LegTrace, Pokemon_LegReplay and leg_trace.py now come from #464; this PR makes RouteLogLines emit one [Leg] record per RouteStep that says how the end was known; with no movement measurement, hold ends are labelled dead_reckoned, not wall or seen. A new fault file ConsoleFailures/12.03-...-pallet-shore-east-unmoved.json replays that case in the overworld sim (new moved_less fault); it is red today (the walker finishes "arrived" while the player is still at (6, 10)), so it is wrapped as known-failing and flips once the walker measures movement. Tested with the headless runner: LegTrace, RouteLog, OverworldRecovery, OverworldSim and the new replay suite pass; test_leg_trace passes in Python.

🤖 Generated with Claude Code

https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss

dayfine and others added 8 commits November 12, 2026 23:07
…ilures on the overworld sim (12.03 pallet_shore)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
The walker measures whether the map scrolled during each plain walk. run_route() takes a DONE walk with no movement, where the plan expects tiles (known start tile, or the fewest from the tiles a first hold may start from), as STOPPED at its start: it is walked again from there, and fails if it moves nothing twice. A hold that may have been at its wall already is a bump and is taken as walked. 12.03 pallet_shore replay case no longer known_failing.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…T-menu test expects the failure

A plan without a start names holds from tiles it does not give, so a hold that moved nothing there is not judged. The START-menu-open sim test now sees the unchecked walk fail as moved nothing, which it could not before.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
… sim); the BDSP sim rejects it

Pokemon_LegReplay is owned here; the kind is added so the FRLG leg-records PR (#451) carries an identical copy and the two merge without a conflict in this file.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 3 commits November 13, 2026 04:02
…test (red)

Console B 2026-11-13 02:42/02:46, 12.03 pallet_shore: the first hold after the
Fly walked nothing while Pallet's town banner slid in over the top quarter of
the screen, which is all the movement check compared ("changed by 223.0"), so
the hold was taken as walked and every later tile was 2 rows off.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…oth changed

A town's banner over the top quarter is no step (console B 2026-11-13 02:42/02:46,
12.03 pallet_shore: "changed by 223.0" on a hold that walked nothing).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
…rlg-stage8

Conflicts resolved keeping both: the bike-tap counted legs, gates and leg breakpoints of the lab branch, and #451's [Leg] records, movement check and console-failure replay.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
…arty menu, #463 surf legs that land

Conflict: OverworldRecovery.cpp's DONE step; took #463's after_done() and kept
#451's m_known update inside it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
Conflicts: the hold's movement check (walk_hold) goes into stage20's
walk_hold_step after its turn and fade watch; tap-counted legs stay
unmeasured, on the console and in the sim; the [Leg] trace's end lines
keep stage20's leg screenshot; RoutePlanning keeps gates and pause and
gains the checkpoint; RouteRun keeps gates, walk-on starts and the leg
breakpoints, and m_known is kept up in after_done().

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
20.01 to 20.04's scene end pose sweeps: the room's walk leaves the player
on (6, 6), a tile past the table's (6, 7), and before_lorelei's one leg,
north 1 onto (6, 6) under Lorelei, only bumps. #451's movement check
walks it again from (6, 7) and fails "moved nothing twice".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
…e there

Verdict, 20.01 to 20.04 scene end pose (enter_*_room ends on the tile
under the Elite Four member, lances_corridor likewise): the movement
check is too strict here, not a walker or plan bug. The room's OnFrame
walk is a scene the sim doesn't play; the sweep leaves the player one
tile on from the table's pose, on the route's target itself, so the
route's only leg (north 1, a person straight ahead) bumps and moves
nothing. The player is where the route ends. Walking it again from the
table's tile moves nothing again, and the route failed.

Narrowed: for a route of one tile from the caller's start, onto a tile
with a wall or a person straight ahead, expected_tiles() is 0, as for a
hold that may start at its wall already. The route is over either way;
if the hold was lost instead, the talk after it fails visibly, as
before #451. Any longer leg, or one with more of the route after it, is
still checked (12.03 pallet_shore's east 4 stays caught).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
Verdict, 19.04b below_ladder: a real plan bug, which #451's movement
check exposed. push_onto_switch leaves the boulder on the switch (20, 16),
but the maps still show it where it started, (7, 18), so the planner
walked below_ladder's first leg south 1 "to the wall" (20, 16), onto the
boulder. The hold only bumps it. Before #451 the walker took that leg as
walked and believed the player on (20, 16) while they stood on (20, 15);
the holds to walls after it happened to bring the sim back. With the
check, the leg moved nothing twice and the route failed (red: the
Simulation 19.04b test).

Fixed in the plan, as for the row's other pushes: by_boulder_8_corner
steps west to (19, 15) first, off the boulder's column, and below_ladder
plans from there (west 2, south 3 to (17, 18), then as before). The
golden's 19.04b lines follow.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
…/dev-frlg-stage8

#451's leg records and movement check, plus the 19.04b plan fix and the
narrowed check for one-tile routes to a wall. The walker files take
stage20's versions: the lab's older #451 copy is superseded.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 2 commits November 13, 2026 15:11
The shared [Leg] record writer and reader (LegTracer), the fault-file reader
for replaying a console failure on a sim (LegReplay), their unit tests, and
the Python grader and fault-file tools under devtools/maps. Taken unchanged
from the FRLG walker commit and nav/leg-replay (#447) so both games build on
one copy.

(cherry picked from the shared parts of commit 9499c3f
and devtools/maps/leg_replay.py from 98c0b61)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…eg trace and replay

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 13, 2026
…ared leg trace and replay (#464)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 2 commits November 14, 2026 07:01
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant