ClashGO v0.5.0-beta

Added

• Deploy budget — the deploy phase is now wall-clock bounded
  (internal/attack/attack.go). StartDeployBudget arms a deadline
  (~2× the longest legitimately-observed full deployment) and every
  reconcile / sweep / verify loop consults DeployBudgetExhausted()
  between rounds. Observed live before this: a spent Earthquake card whose
  cooldown icon never reads "empty" kept the sweep re-firing taps for
  ~2 minutes after the battle had ended. Leftover slots are now reported
  as undeployed so the attack report shows the partial failure instead of
  a phantom success.
• end_at_percent per-strategy auto-end (pkg/strategy/yaml_parser.go,
  internal/attack/attack.go) — a strategy can declare the destruction
  percentage at which the win is already secured (e.g. valk_spam.yaml
  ends at 50%). In threshold mode the stall timer is deliberately disabled
  below the threshold so a stall can never abandon the win early, and the
  battle only ends when the red End Battle button is actually on screen.
• Battle-outcome correctness (internal/attack/attack.go):
  LastDestructionPercent latches the highest destruction read during the
  battle (monotonic, so a transient 0 as the overlay renders can't win),
  ThDestroyed latches the golden "Town Hall destroyed" banner from an
  optional stall_config.th_banner_zone, and StarsFromOutcome scores
  stars from the outcome (≥50% = 1★, TH = +1★, 100% = 3★) when the result
  panel's OCR fails. ResetBattleOutcome clears both at the top of
  WaitForBattleEndCtx — live, battle 1's garbage 381% read was still
  latched when battle 2's result was parsed, turning a 32% defeat into
  "3 stars".
• StateConnectionLost + StateConfirmExit (internal/game/types.go,
  internal/game/classifier.go, internal/bot/bot.go) — the disconnect
  dialog (TRY AGAIN / RETURN HOME) was misread as StateBattleEnd, so the
  bot tapped result-screen coordinates forever (live: 18:26 boot → 18:28
  ReturnHome fallback → 18:33 restart loop). TRY AGAIN reconnects in
  place; the quit-confirm dialog is dismissed with CANCEL.
• Saved-army slot selection (cmd/… internal/bot/bot.go,
  pkg/strategy/yaml_parser.go) — strategies declare army_slot and
  selectArmySlot clicks that recipe card in the army list (measured
  geometry: first card at ref-y 227, ~54px stacking). valk_spam.yaml
  targets the 4th recipe.
• Inter-attack cooldown (min_seconds_between_attacks, default 30s)
  — without it the bot re-attacked ~8s after every Return Home with
  whatever the camps held (live: three near-identical defeats in under
  four minutes).
• BlueStacks adbd wedge workaround (internal/adb/transport.go,
  internal/adb/shellpipe.go). In the wedged state adbd answers every
  bare command with FAIL "closed" while compound commands
  (getprop x; <cmd>) run normally. shellWedgeFallback rewrites the
  service string, stripPrefixLine drops the extra output line, and the
  exec-service path falls back to shell: (with a 16→12-byte screencap
  header normalizer) — the wedge survives adb kill-server, so this is
  the only durable fix at the transport layer.
• screenSize recovery ladder (internal/bot/bootorchestrator.go) —
  a failed wm size now escalates retry-transport → adb kill-server +
  start-server between attempts instead of burning the budget against the
  same wedged socket.
• Phase-level deploy offset propagation (internal/attack/deploy_planner.go,
  pkg/strategy/yaml_parser.go) — a phase pin like offset: 130 # Deeper in for EQs now applies to every unit in the phase (per-unit offset
  still wins).
• Post-deploy spell reconciliation (internal/attack/spell_deployer.go)
  — Amount: "All" resolves to the card's live OCR count (11 rage on one
  card, 4 EQs on another) instead of a hardcoded 5, and
  VerifyAndReconcile re-reads the slot afterwards and re-fires the
  difference. A no-progress stop (two identical positive reads) and a
  zero-OCR-streak bound stop a spent card from eating taps forever.
• cmd/wedge_probe — standalone live harness for the adbd-wedge
  detection + compound-command fallback.
• Test coverage for every new decision path — deploy_budget_test.go,
  battle_end_test.go (destruction → stars, trophy/star pixel anchors,
  defeat fixtures), spell_deployer_test.go (tap-level assertions + formula
  geometry), slot_manager_test.go (duplicate manual-label hijack),
  testharness_test.go, bootorchestrator_test.go (wm-size fallback),
  classifier_stuck_test.go, formula_test.go, yaml_parser_test.go, plus
  Python strategy-contract tests (tests/test_spell_placement.py,
  tests/test_valk_strategy.py) wired into make test.

Changed

• Loot OCR is content-located, not slot-located
  (internal/game/loot.go) — loot rows are found by detecting bright,
  low-saturation connected components (the white/gold digits) inside a
  generous search zone and grouping them into horizontal text lines, so the
  reading no longer depends on fixed HUD geometry. Digit ROIs get relaxed
  padding for narrow columns (the League Bonus gold was read 10× low when
  the right edge clipped the trailing digit), and implausible reads are
  rejected instead of being fed to the stall timer.
• Frame loop 10 Hz → 4 Hz (internal/bot/bot.go) — during a battle the
  deploy and battle-end goroutines run their own captures at their own
  cadence, so the full-screen classify at 10 Hz was mostly redundant
  (~55ms captures + classify pinned a core). 4 Hz still catches the result
  overlay fast enough for ReturnHome and the stuck-watchdog.
• ArmyCamp rule hardened (internal/game/classifier.go) — the lone
  brown pixel anchor also passed on ordinary village frames (live:
  (479,149) = RGB(61,53,62)), which made processFrame press Back on the
  real village and open the quit-confirm dialog the bot then sat on for the
  5-minute boot grace; the rule now requires both army-overview tab-header
  anchors, and a frame that still shows the real Attack! button is treated
  as the main village and never Back-pressed.
• Battle-search / bonus ROIs widened (assets/battle_loot_rois.json) to
  match the new content-located loot reader.
• make test now runs the full contract suite — go vet, Go tests
  (with the local OpenCV LC_RPATH injected via pkg-config), and the
  Python strategy tests, with a graceful skip when no pytest-capable
  interpreter is present.
• valk_spam.yaml declares end_at_percent: 50 and army_slot: 4.

Fixed

• Fabricated 3-star victories — the destruction-rule read used to
  OVERRIDE the result panel's parsed stars, so a garbage stall-ROI percent
  (381%, 851%, 991%) turned live defeats into wins. The visual parse is now
  ground truth and the rule read is only a fallback when the panel parse
  fails outright.
• Star art read as victory on the defeat screen — the result-panel star
  parse is now two-pass and defeat-safe; pinned by the new
  internal/game/testdata/screen_defeat.png fixture.
• Stale manual_labels.json entries creating duplicate unit identities
  (internal/attack/slot_manager.go) — a manual label now loses to a
  template match on a different slot. Live, a stale "grand warden" label
  on the unidentified Dragon Duke slot stole the warden's main-phase deploy
  and the Duke was never swept (it looked deployed).
• wm size wedge aborting the boot — falls back to the live
  screencap's header for authoritative pixel dimensions (this is the same
  fix family as 0.4.0's boot fallback, now with the retry ladder above).
• Pipe read deadline left armed after service negotiation
  (internal/adb/shellpipe.go) — the idle discardReader inherited the
  shell-timeout deadline and marked a perfectly healthy pipe broken
  ("adb shell pipe reader stopped: i/o timeout") mid battle-end wait.

Verified

• Full live session on BlueStacks Air (860×732, SM-G998B profile): boot →
  classify → attack search → loot OCR → EDrag Rush and Valkyrie/EQ runs →
  result parse (stars derived from destruction, defeat fixture matched) →
  graceful shutdown.
• make test green: go vet clean, all Go packages pass, 24/24 Python
  strategy-contract tests pass.