fix(goal): kernel correctness, injection consolidation, and hardening

[LeXwDeX]

Summary

Fixes two verified correctness bugs, consolidates the prompt injection design, adds two hardening guards, and establishes kernel test coverage from zero (56 tests). Tracked via OpenSpec change goal-loop-correctness.

Bug fixes (verified against code, TDD-exposed)

turns_used done-state inflation

turns_used was incremented on state transitions (active→done), not just on continuation dispatches. Done is a state label, not a turn. This was the root cause of the agent self-declared completion double-count (model calls goal(complete) mid-loop → markDone added a second +1 on top of updateAfterJudge's prior +1).

Fix: increment ONLY in updateAfterJudge continue branch (immediately before continuation dispatch). Removed +1 from: updateAfterJudge done branch, markDone.

goal.updated(done) double-publish

updateAfterJudge done branch AND deleteAndPublishDone both published the same terminal goal.updated(done) event. Violated the documented bus contract (updated(done) → cleared); risked double-firing hooks keyed on the terminal event.

Fix:updateAfterJudge done branch no longer publishes. deleteAndPublishDone is the sole owner of the terminal event sequence.

Design: prompt injection consolidation

The model received goal context via 4 fragmented channels (static system prompt, noReply progress message, ignored continuation prompt, tool description commanding per-turn status check). The model was forced into a redundant goal(status) round-trip every turn.

Discovery: the old continuation prompt used ignored: true, which is filtered from BOTH the transcript AND the model's messages (message-v2.ts:208). The continuation instruction never actually reached the model — the loop continued only because prompt() created a new turn.

Fix:

• SystemPrompt.goal(sessionID) now yields live goal state dynamically (was static text; the Effect shell already existed)
• continuation prompt + progress message merged into a single injection with plain text part (no ignored) — now actually reaches the model and is visible in transcript
• goal(status) tool reframed from mandatory per-turn to optional steering check-in

Hardening

• Transport errors count toward pause budget: judge transport failures (timeout/network) previously returned parseFailed: false, resetting the failure counter. A flaky provider could alternate bad-JSON and timeout indefinitely without triggering auto-pause. Now returns parseFailed: true.
• Zombie-goal freshness guard: if a goal is active with turns_used: 0 and created_at > 120s with no assistant message (initial kick failed), auto-pauses instead of orphaning silently.

Test coverage

File	Tests	Coverage
test/goal/judge.test.ts	13	parseJudgeResponse 4-step fallback, transport error handling
test/goal/loop.test.ts	14	shouldPreempt edge cases, isStaleZombie predicate
test/goal/goal.test.ts	18	state machine (set/pause/resume/clear/markDone), fiber discipline, event contract
test/goal/prompts.test.ts	11	renderContinuation merged output, renderGoalSystemBlock
Total	56	was zero

Verification

• bun run typecheck: green (0 errors)
• bun test test/goal/: 56 pass, 0 fail
• Pre-push hook: turbo typecheck 29/29 FULL TURBO

Remaining (live e2e)

tasks.md §11.1-11.3 (live /goal lifecycle in tmux, SSE event stream validation, model behavior observation) deferred to reviewer/user validation with an actual provider.

OpenSpec

Change: goal-loop-correctness (proposal + design + specs + tasks all validated --strict)