fix(goal): stop cancels and errors from re-running the goal, drop a stray edit

[ericleepi314]

main is red. CI on #779: 1 failed, 9474 passed — tests/server/test_goal_control.py::TestMaybeContinueGoal::test_cancelled_and_error_turns_skip_judging. Two defects, both introduced by #779.

1. A user cancel or provider error restarted the goal loop

#779 made /goal auto-continue past a turn the agent loop cut short, keyed on subtype not in ("", "success"). But _run_turn also returns cancelled on AbortError and error on an exception — both got swept in.

So pressing ESC during a /goal loop re-enqueued the work you had just killed (the _inbox.empty() preflight doesn't help — an interrupt leaves it empty), and a provider 5xx/auth failure retried up to the goal's whole turn budget.

It contradicted the function's own docstring ("Skips when: … the turn was cancelled/errored") and _after_wakeup_turn's rationale, which cites the goal loop's no-continuation-on-error guard as its precedent. A test already pinned this; #779 broke it.

Narrowed to EARLY_STOP_SUBTYPES.values() — the stops the agent loop produced. Those mean "the model did not finish", so retrying is right and matches pre-#779 behaviour (the turn arrived as "success", was judged not-done, and the loop retried). cancelled/error are the same category as the hook stops deliberately excluded from that map: the user or the provider ended the turn, not the harness cutting it short.

2. A stray edit reached main

agent_server.py carried, verbatim:

if early_stop:  # MUTANT: short-circuit the budget tick
    should_continue = True
    continuation = continuation or "[Continuing toward your standing goal]"

Mutation-testing scaffolding committed by accident in #779 — the tree was mutated in place while the commit was taken, and the staged diff wasn't read.

Measured impact, because the comment overstates it: apply_verdict still runs before this block, so turns_used still ticks and the goal still deactivates at its cap. At max_turns=3 it yields 3 continuations instead of 2 — one extra model turn past budget, bounded, not a runaway. Still wrong: it overrides the goal's own decision to stop.

Tests

The cap test added in #779 cannot catch (2) — the cap genuinely holds under it, which is exactly why it slipped through. So the new test asserts authority instead: patch apply_verdict to say stop, assert nothing is enqueued. It fails against the exact code that shipped.

Also adds a cancel/error guard test in this file — the pre-existing one lives in test_goal_control.py, which is why the targeted runs during #779 never saw it.

Both new tests carry positive controls: this harness swallows exceptions, so a bare "nothing was enqueued" assertion would otherwise pass for the wrong reason if a stub were missing.

Verification

• Full suite back to the local MCP-only baseline; test_goal_control.py no longer among the failures.
• tests/server: 469 passed.
• Both defects mutation-tested by re-introducing the exact shipped code.
• Staged diff read before committing, and git show HEAD confirmed free of scaffolding.

🤖 Generated with Claude Code

[github-actions]

Test Results

    1 files      1 suites   8m 3s ⏱️
9 484 tests 9 478 ✅ 6 💤 0 ❌
9 752 runs  9 746 ✅ 6 💤 0 ❌

Results for commit 30f18b8.