Skip to content

PLZA 03.08-03.11: port plza/autostory's console-passed steps - #623

Closed
dayfine wants to merge 40 commits into
plza/backup-retryfrom
plza/port-original-0308-0311
Closed

dayfine wants to merge 40 commits into
plza/backup-retryfrom
plza/port-original-0308-0311

Conversation

@dayfine

@dayfine dayfine commented Nov 16, 2026 •

Copy link
Copy Markdown
Owner

Depends on #622 (base plza/backup-retry; the diff shows only this change).

The lean chain's 03.08–03.11 had drifted from plza/autostory, whose versions passed on console B on 2026-11-14 while the lean ones went 0 for 10 or never ran (audit: ~/BDSPLab/notes/plza-divergence-2026-11-16.md). This ports the original's behaviour onto the lean engine. 03.08 keeps #622's measured walk from Taunie's rocks to 100, then runs the original patrol: lock-on brackets, the nearest plate not caught, throw again after a miss, the party bar ending it, and the banner/plate tally when the bar is unread. ZoneSearch is dropped. 03.08b uses the pillar-steered gate walk and A Leave. 03.09 gets five_registered back, 03.10's choose_1 presses A through the floor question, and 03.11's leave_mable ends on control. The "Registered!" banner is watched over the throw plus 2 s with run_until, not the original's CoveredWatch. The PR also fixes the 4 AutoStoryBackups rewind tests, whose stand-in catch_in_zone the registry refused as console_only.

On top: the lab's 03.08 zone-start commits (a timed 5.6 s hold from Taunie's rocks to 100, then the patrol even with no read), and the patrol's walks between 100 and 60 now use plza/autostory's walk (walk_to_road_stop, RoadWalkReads): it steps on after every look and needs a read only to arrive, where the lean marker walker stood, nudged and sidestepped on unread looks and ended at -1, 74 and 92 short of 60 on console B (2026-11-16). The lean walker stays for walk_to_marker. New suites: ZoneStart, RoadWalkReads; the stale 03.08b leave-gate case now runs A Leave to the party bar.

Tested: the touched files compile locally. The full build and the PLZA unit tests are left to Fork Draft CI, because a console run blocked the local build. Ported suites: CatchPlan, CatchStop, GateWalk, ZoneGateReader, LockOnBracketDetector. No console run yet.

03.08: after ZoneStart's measured walk from Taunie's rocks to 100, the
original patrol (100/60 back and forth, a camera sweep for the lock-on
brackets, ZL on the nearest plate not caught, throw again after a miss
that locks on again, weaken a target that broke out, fight off caught
species), the party bar ending it and the banner/plate tally deciding
when unread. Ports LockOnBracketDetector, DexRegisteredDetector (the
banner watched over the throw and 2 s after, not CoveredWatch),
CatchPlan, CatchStop, BracketWatch, SpeciesCheck and Zone1Road; drops
ZoneSearch. walk_back_to_zone_road ends facing the marker.

03.08b: the pillar-steered gate walk (walk_to_zone_gate_entrance,
ZoneGateReader, GateWalk) at heading 265 and A Leave as a scene step,
replacing run_to_gate there and leave_zone_gate.

03.09: five_registered back as its own step that control ends.
03.10: choose_N presses A through a question's text when no choice
shows, as the original did; floor_question dropped.
03.11: leave_mable ends on control.

Fixes the AutoStoryBackups rewind tests: the stand-in catch_in_zone is
no longer console_only, so the registry runs it on the script backend.
The lean marker walker stood on every look with no distance taken, nudged
the camera and sidestepped through lanes, then ended blocked: on console B
(2026-11-16 16:17-16:21) the walks toward 60 ended at no read, 74 and 92,
sideways, and Mareep at 60 was never searched for. plza/autostory's
walk_to_objective() steps on after every look and needs a read only to
arrive; its first read is taken at once and a misread far from the last is
dropped (three agreeing far reads replace a wrong last).

walk_to_stop() (03.08's patrol and 03.08b's walk back to the road) now uses
walk_to_road_stop() with RoadWalkReads, that rule plus the road floor for a
first read ("1", "21" for 101). The lean walker stays for walk_to_marker()
and read_marker_distance().

Tests: RoadWalkReads, and ZoneStart's cases back (dropped with ZoneSearch's
tests), now with no read after the timed walk going on to the patrol. The
stale "03.08b's way out of the gate needs a console" case now runs A Leave
through the fade to the party bar.
…e full map

read_map_pin() finds the player's arrow on the full map (PLUS) in Wild
Zone 1's window. 03.08's walks (rocks to the 100 spot, between 100 and 60,
back to 60) become legs to map waypoints (Zone1Map, all provisional):
face the bearing by the minimap, hold forward for the distance's time, read
the pin, correct twice at most, then stop as "leg_off_target". The marker
distance is read only as a logged cross-check. CI pins the
CommandLineTests commit with the two map frames.
… after 3 misses

The plate's height on screen is the distance signal: console B's first-throw
catches had it at 0.49-0.50, the Mareep missed 11 times at 60 at 0.40 (plate
width and bracket height did not separate them). Below 0.45 the catch steps
toward the target (4 steps before the first throw, 2 after each miss); after
3 misses at one lock the stop ends. Each throw logs its plate height and
hit/miss.
The scene machine's wait on a still screen looks until the screen changes, so
the first party bar after it is the wait's look and the machine's own party
bar hold starts on the next one, as in 03.02b's plaza scene. The test gave one
party bar before the repeated last look, so the hold never reached a second.
…bjective

On console B (2026-11-16 17:34) 03.09's walk from Wild Zone 1's far gate read
no distance ("t", "", "1"), stepped on blind and passed on the first A
prompt it met on the street, 4 s in: the walk took any prompt or dialog as the
objective's, and so did its end check. plza/autostory's walk had the same rule
but passed because its start read 37 and it walked to 13 before Taunie's scene.

MarkerArrival now decides arrival: a prompt or dialog counts only near the
objective, by a distance read taken (agreed reads, MarkerDistance) at 20 or
less; farther, it is walked past. Up the lab's steps the digits never read, so
that row uses walk_to_unread_marker: no read over 20 at all. The end check
after the marker goes (a scene at the objective) waits only when near.
…teps

Keeps PokemonLZA_AutoStoryGame.cpp at its 1000-line limit: the proof is
chosen in MarkerWalk, not by lambdas in the step table.
…ture

The latest save after 03.08 was the fifth catch's autosave, wherever in
Wild Zone 1 that catch was, and a save on the street past 03.09 read "in a
wild zone" (lab-B 2026-11-16 18:34): 03.08 ran as a no-op and 03.08b walked
from the street.

- save_from_menu (X, Save, B back to the party bar) counts as the
  checkpoint's save. Added at the end of 03.08b (out of the far gate,
  facing its prompt), 03.09 (lab lobby, autosave never seen) and 03.11
  (Mable's office; the game autosaves mid-checkpoint in mable_talk).
- StartSignature gets party (the party bar's filled slots). Signatures for
  03.08 (party 3), 03.08b, 03.09, 03.10 and 03.11.
Console B, 2026-11-16 18:47: 03.09's walk to the lab gate took junk OCR
("E5", "v1", "E6": 5, 1 and 6 agreed) for "6 to go" and passed on a street
dialog; the gate's real scene (marker gone at 12, Ryon's dialog 10 s later)
then played inside up_the_lab_steps and lab_lobby, and 03.10 met Ryon's
"cinque specie" dialog at the doors.

- MarkerArrival: a read below 10 is taken only after one of 10 or more; a
  prompt or dialog with the marker still up counts only below 10 (or with
  no read); with the marker gone (the scene took it) up to 20 as before.
  The plza/autostory passes (37 down to 13 and to 10, marker gone, then the
  dialog) still arrive.
- The marker walk presses B on a dialog that is not the objective's.
- mash_to_move_hints takes "scene_start_ms" (SceneMachineConfig::
  start_window): 03.09's lab_lobby fails with no scene in 3 s instead of
  passing on control. five_registered keeps passing at once: the original
  pass had control there in 1.2 s (2026-11-14 21:01:47).
- 03.09's whys name Ryon. CLT frame of his dialog (dayfine/CommandLineTests
  plza/0309-gate-frames, pinned in CI) read as a dialog, not a prompt.
- choose_<index>'s step moves from AutoStoryGame.cpp to the Scenes module
  (choose_row()), unchanged, bringing the file back under 1000 lines.
Console B, 2026-11-16 18:44:57: 03.08b zone_1_to_far_gate failed as
gate_not_seen (no gate on 13 looks in 3000 ms) while a wild Mareep had
engaged at the zone's pillar. walk_to_zone_gate_entrance() now walks on at
the last angle through looks with the battle HUD, counting them toward
neither the first-seen window nor the timeout, and fails as "wild_battle"
after 20 s still in battle. CLT frame of the battle (pinned in CI) reads as
a battle.
…ter catches

Console B, 2026-11-16 18:43: the leg to the 60 spot ended at (384, 253)
facing 245, next to the pin (386, 252); the Mareep's brackets showed at
x 0.69 (0.081 tall, plate y 0.31), and 4 steps of 0.5 s toward it brought
the plate to y 0.45 before the lock.

The arithmetic: x 0.69 is 0.19 of the width right of the middle, 11 deg
with the view taken as 60 wide (VIEW_DEGREES' guess), so the target bears
256. The leg covered 49.8 map pixels in its 7104 ms hold, 7.0 px/s; the 2 s
of steps are 14 px on heading 256: (-13.6, +3.4) from (384, 253), so the
pin goes to (370, 256), on the road's middle line. Still [provisional].

This is ahead and right of the facing (the user's "move more to the right
during the patrol", the Mareep on the right of the screen), which on the
map is west, not east: the patrol walks west at heading 248.

catch_at_stop() now reads and logs the map pin after a catch, out of
battle, so the spots can be calibrated from runs.
Console B, 2026-11-16 18:39: 03.07's enter_zone_1 (mash_to_move_hints)
showed scene_moving from 1.5 s to 11.7 s with no press while the
"received 10 Poke Balls" box was up; there is no frame of the box yet to
build a detector from. SceneMachineConfig::gap_press: once a scene is under
way, one blind press when nothing busy and no sign of control has shown for
that long, and again only after the next cue. Opt-in as mash_to_move_hints'
"gap_press_ms"; 03.07's enter_zone_1 sets 2000. Scripted runs check the one
A, and none without the option.
…tle; a still pin re-aims

Console B, 2026-11-16 19:24 (run 3): the lead's panel shows all through
Wild Zone 1, so BATTLE_STARTED fired on the far pillar with no battle and
the gate walk walked on 20 s into the pillar.

- BattleMarkersDetector: the five markers right of the panel's gender
  icon (one blue triangle, white dots) show only in a battle. A battle by
  the panel is now panel and markers. Two zone frames with the panel up
  and no battle are negative frame cases (CommandLineTests 7a024c6,
  branch plza/0308b-pillar-frame; CI pin bumped).
- The gate walk is done out of the zone (the lead's panel gone on 2 looks
  after it showed), prompt or not, whatever a battle reading says (the
  user). A battle only matters inside the zone.
- Walking on reads the map pin every 2 s; a pin that moved under 2 map
  pixels is a wall: face the heading again and steer by the gate.
- The 60 spot goes back to (386, 252): run 3 saw no new species at
  (370, 256). 03.08b's gate walk passed from there at 17:34 and 18:46.
The menu save hung 2 minutes on console B (2026-11-16 20:05): a save
file's first X opens the menu under a "Save the game" tutorial card, and
upstream's save_game_to_menu() waited on it with no log. save_from_menu
now reads the screen each look and presses one button per held screen:
X, A on the tutorial card, R on the main menu (the save is R, not a menu
entry), A at the confirmation, B through the saved text and out, with a
log line per screen and a 30 s bound.

03.08b drops face_zone_1_gate (a hand-written leg) and saves where
leave_zone_1 ends; 03.09's start signature is party 5 out of the wild
zone, with no prompt field. CLT: the tutorial card and the menu after it
(dayfine/CommandLineTests plza/menu-save-frames).
The plan loader refuses a start signature that names two checkpoints:
03.11's (party 5, not in a wild zone) was 03.09's too, so the program
stopped before the first event. 03.10's matched the street as well.
Both rows go back to unchecked until a lab lobby / 3F frame gives them
a reading of their own.
03.11 leave_mable hung 60 s on Mable's mission/reward screens as
mash_to_move_hints, which presses nothing on screens no detector reads
(console B 2026-11-16 20:42). plza/autostory ran it as a bare "scene"
and passed (2026-11-14 22:17). Port that step's rule into the scene
machine as action "mash", the default of a "scene" step with no action.

Copied from origin/plza/autostory:
- dispatch: MASH name != "a" -> B, CONTROL_ENDS (PokemonLZA_AutoStoryGame.cpp:607-612)
- 120 s timeout (Actions/PokemonLZA_AutoStoryScenes.cpp:108)
- B at a box, A at a choice, a fade waits, the party bar waits, B on every
  other look (PokemonLZA_AutoStoryScreenEvents.cpp:66-81)
- A at a tutorial card (Actions/PokemonLZA_AutoStoryScenes.cpp:84-88)
- one press (160 ms / 240 ms) per look, no extra wait (Actions/PokemonLZA_AutoStoryScenes.cpp:39-49)
- done on the party bar held 1 s, at once if from the start
  (OVERWORLD_HOLD, PokemonLZA_AutoStoryScreenEvents.h:81; TalkOver calm,
  Pokemon/AutoStory/Pokemon_SceneJudges.cpp:52-67; CONTROL_ENDS,
  Actions/PokemonLZA_AutoStoryScenes.cpp:99)
Not ported: its 45 s SCENE_START_WINDOW failure (Scenes.cpp:91-97); the
lean machine's A-prompt-held-1 s end stays.

SceneMachineConfig gets blind_press_every (default the old 2 s
BLIND_PRESS_EVERY); "mash" sets blind_press_after 0, every 0, button B.
leave_mable, five_registered and arrive_3f go back to bare scenes as in
plza/autostory's plans; the lead-in lint now applies only to "a" scenes,
since control ends a bare one at once.
Mable's office in 03.11 (her talk autosave) shows a talk prompt outside a
wild zone, which matched 03.06's start, so a run asked for 03.08 ran
03.06's legs in the lab (console B, 2026-11-16 21:05). At the Center
counter the party is the starter alone (1 registered on the backup card;
03.08's party of 3 is the starter and the tutorial's two catches), so the
signature carries party 1 and the office's party of 5 no longer matches.
…runs it

An earlier start matching the screen is no sign of being there: 03.09 to
03.11 all read as party 5 outside a wild zone, so a run asked for 03.11
from Mable's talk autosave ran 03.09's legs in her office (console B,
2026-11-16 21:10). Only a signed start is checked; rewinds still place a
backup card by checkpoint_at().
Their menu saves (save_from_menu) leave each a save the run checks, so the
built-in story's row list expects ':autosave' for them.
Reads SerialPrograms.log files for the autosave badge per checkpoint run and
lists the unverified plan rows a passed run saw autosaving in their last step.
--apply rewrites those rows to "save": "autosave" with autosave_step and
cites the log in their source text. A badge before the last step, in a failed
run, or a passed run without one leaves the row unverified.
A chained run with "Developer: skip forced reloads" on goes on from the
live screen into a row marked reload_when_chained "always" and logs
"[Checkpoint] <id>: reload skipped (developer option)". Default off.
run_plza_segments() takes SegmentRunOptions (attempts per checkpoint and
the forced-reload switch); RewindPlan carries them.
Run 9 of 03.09 (console B, 2026-11-16 21:33) took 4 5.4 s after 40 at
Wild Zone 1's exit (the marker showed 39) and passed on the exit pillar's
prompt. A read of 20 or less is now taken only if the walk can have come
that near since the last read taken (5 m/s plus 2 m of OCR jitter, from
run 10's 23 down to 12 in 3.2 s); rejected reads are logged.
French's 01.02 choice panel starts at x 1198 and its arrow at x 1185, left
of the box's x 1256, so CHOICE_BOX never fired (run 11). The box now starts
at x 960. The rows reader already scans the highlight's left end; the French
frame checks both. CI's CommandLineTests pin moves to plza/choice-fr.
PLZA AutoStory: choice arrow box covers French's wider panel
@dayfine

dayfine commented Nov 17, 2026

Copy link
Copy Markdown
Owner Author

Superseded by #644 (one lean branch; split for review later)

@dayfine dayfine closed this Nov 17, 2026
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