fix(goal): spend the budget per judged turn and never leave a goal active on an idle session

[filipeforattini]

Problem

Four defects in goal mode (/goal) left the budget meaningless or the goal stranded:

1. Budget unit. GoalRuntime.beginTurn ran before every provider attempt — each tool round-trip and each retry — and charged one turn each time, so the default of 20 was exhausted by a goal that read fifteen files and ran five commands. Meanwhile the goal block said "Turn 3 of 20", the continuation said "turn N of M", the TUI toast said "20 turns" and the budget dialog said "provider attempt budget".
2. CAS failure ended the turn silently. /goal-budget and /goal-resume bump updated; if that landed between the fresh read in prompt.ts and the compare in afterTurn, afterTurn returned undefined, the loop broke, nothing was recorded and the goal stayed active on an idle session.
3. Step ceiling and stall left the goal active with no trace. The step-budget stop broke out of the loop before afterTurn; the stall watchdog called state.cancel rather than SessionPrompt.cancel, bypassing goals.pause.
4. Evidence and truncation. Observed tool output was collected since goal.created, so one read on turn 1 satisfied evidence.length > 0 forever; and [...gates, ...observed].join().slice(-24000) cut gate results first.

Fix

• A turn is one full agent turn ending in a judge cycle. turns.used is incremented in SessionGoal.apply (once per judged turn, never on WAIT); beginTurn is an admission gate only. decide treats the turn being judged as spending one, so the budget check is used + 1 >= max. The judge prompt, budgetReason, the TUI budget dialog, the resume toast and specs/goal-modes.md all say "turns". Default stays 20; runaway protection remains the retry policy, loop guard, step ceiling and MAX_JUDGE_FAILURES.
• afterTurn takes the decision on a record and writes it only if that record is still stored. On a lost race it re-reads and decides again on the fresh record (a raised budget can turn a stop back into a continue); a second loss logs and pauses the goal with a reason.
• The step-ceiling stop and the stall watchdog both pause the goal through goals.pause with their own reason (stopped at the step ceiling after N steps; ..., stalled: no output for ...) before ending the turn.
• Evidence is scoped to the current turn: everything since the last spent turn (goal.judged, new optional field, falling back to goal.created). Anchoring on the last spent turn rather than the last user message keeps a turn parked on background work and the report that re-enters it as one turn. SessionGoal.evidence puts gate results last so the tail cut keeps them.

Test

Regression tests fail on main and pass here (run with the source stashed and the new tests in place, then again after restoring):

• test/session/goal.test.ts: budget check (used: 2, max: 3 is the last turn), apply spends one per judged turn and none on WAIT and stamps judged, evidence keeps gate results through truncation — 3 fail before (SessionGoal.evidence is not a function, wrong counts), 20/20 pass after.
• test/session/prompt.test.ts: retry inside a turn spends nothing (before: retry refused, 1 hit); many tool steps are one judged turn (before: paused after the first attempt); terminal failure spends 0; background WAIT spends 0; /goal-budget landing during judgement is recorded against the fresh record (before: active); step ceiling pauses the goal (before: active); stall pauses the goal (before: active); a tool result from an earlier turn is not evidence for a later claim.
• Whole test/session directory green; bun run typecheck clean in packages/redcode and packages/tui.

Not done: packages/sdk/openapi.json / types.gen.ts not regenerated for the optional judged field (the build script regenerates the whole client; no client reads the field). Spec line about SessionV2 charging per provider attempt left as is — that is a different runtime in packages/core. The uncommitted .changeset/repair-goal-attempts.md in the main checkout says retries count against budgets; this PR supersedes that behaviour and does not touch that file.

https://claude.ai/code/session_01KAcoL6wpgVEs2ebUrUGpHS

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[blacksmith-sh]

Found 2 test failures on Blacksmith runners:

Failures

Test	View Logs
e2e\regression\session-rename.spec.ts/ renames and closes the session tab from its context menu	View Logs
e2e\user-story\model-selection-flow.spec.ts/ creates a session in a new project, connects OpenCode Go, and selects its model	View Logs

^{Need help on this PR? Tag @codesmith-bot with what you need.}