fix(lua.endpoints): dismiss win overlay so endless mode stays responsive

[S1M0N38]

Problem

After winning the ante-8 boss blind, play() returned won=true the instant the win was detected — but win_game() fires right after and raises the win overlay (G.OVERLAY_MENU) while pausing the game (G.SETTINGS.paused = true). The bot was left in a paused session where every subsequent endless-mode endpoint ran on wall-clock REAL time instead of turbo, leaving endless play permanently degraded (~9s per play vs ~1.6s).

Root cause

Two compounding issues in src/lua/endpoints/play.lua:

1. The polling event was killed by the pause. The condition event was created with created_on_pause = true, but Balatro's event.lua ignores that config field — it only honours pause_force (self.created_on_pause = config.pause_force or G.SETTINGS.paused). The event was created while unpaused, so it got pause_skip'd the moment win_game() paused, and never recovered.

2. The overlay was never dismissed. The won branch responded immediately and left the overlay up, keeping the game paused for all following requests.

Fix

In src/lua/endpoints/play.lua:

• Switch the polling event from created_on_pause = true → pause_force = true so it keeps running while paused.
• Dismiss the win overlay inside the event: if G.GAME.won and G.OVERLAY_MENU then G.FUNCS.exit_overlay_menu() end. This is safe because the overlay only appears after ROUND_EVAL is entered (so G.round_eval exists), and the delayed win events (Jimbo speech, endless-round text) guard against a nil G.OVERLAY_MENU.
• The won path now flows through the existing has_blind1 and has_cash_out_button checkpoint before responding, instead of short-circuiting early.

Tests

Two new regression tests, both driven live rather than via save/load fixtures — load resets the run and discards the paused/overlay state, which masked the bug entirely:

• test_play_endless_mode_after_won — drives a full win, continues into endless mode, and asserts the next endless play stays responsive (elapsed < 5s). This is the test that fails without the fix (~9.5s) and passes with it (~1.6s).
• test_won_persists_through_endless_cycle — asserts won=true survives the cash_out → next_round → select cycle.

Removed the state-SELECTING_HAND--won-true fixture from fixtures.json: the save/load mechanism structurally cannot capture the paused/overlay state, so it could not represent this scenario.

Verification

make all green against the documented environment (.envrc):

lint / format / typecheck : clean (63 files unchanged)
tests/cli                 : 157 passed, 14 skipped
tests/lua                 : 513 passed, 2 skipped

⚠️ Reviewer / CI note

The responsiveness test requires the turbo profile + headless render set in .envrc:

export BALATROBOT_RENDER=headless
export BALATROBOT_SETTINGS=turbo

Under the slow default (render=headfull), every game op is ~5x slower and the elapsed < 5s assertion becomes unreliable. There is currently no CI workflow in .github/ setting these, so reviewers should direnv allow (or source .envrc) before running tests/lua locally.

Commits

• a621bab fix(lua.endpoints): dismiss win overlay so endless mode stays responsive
• 282244d test(lua): add endless-mode win overlay regression tests
• 81d2ac7 docs(skill): shorten balatrobot serve startup wait

[Copilot]

Pull request overview

Fixes an endless-mode responsiveness regression caused by the ante-8 win overlay pausing the game and preventing the play endpoint’s polling event from completing, leaving the bot session effectively “stuck” running on wall-clock time.

Changes:

• Update play endpoint polling to run during pause and dismiss the win overlay during ROUND_EVAL so endless mode remains responsive.
• Add two live-driven Lua API regression tests to validate endless-mode responsiveness and won=true persistence through the endless transition cycle.
• Remove an unrepresentable save/load fixture for the paused/overlay win state and shorten the CLI skill doc startup wait.

Reviewed changes

Copilot reviewed 4 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/lua/endpoints/play.lua	Keeps the polling event running while paused and dismisses the win overlay before responding, so endless play doesn’t degrade.
tests/lua/endpoints/test_play.py	Adds a live regression test asserting endless-mode play remains fast after a win.
tests/lua/endpoints/test_gamestate.py	Adds a live regression test asserting won=true persists across the endless round-transition cycle.
tests/fixtures/fixtures.json	Removes a fixture that cannot represent the paused/overlay state via save/load.
.agents/skills/balatrobot/SKILL.md	Reduces the suggested startup sleep before calling the CLI health check.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.