fix(agent): reserve the final step for a terminal response

[plombeer31]

Summary

A run that used every configured step for non-terminal tools ended as stalled with the synthetic (stopped: max_steps reached without a reply) — the model never got a chance to deliver the requested evidence summary even when the work was done (issue #107).

This PR adopts the fix from #269 by @mikemikimike (cherry-picked with authorship preserved — thank you!) and hardens its one gap: cancellation semantics on the reserved final step.

Changes

From #269 (unchanged):

• The last allowed inference (i === maxSteps - 1) becomes a finalization step: the prompt and the native tool payload expose only reply/finish (new StepContext.terminalOnly flag), and a transient notice tells the model to summarize.
• Non-terminal calls on that step are rejected at both parse sites in executeStepInner before any execution, routing through the existing one-shot repair.
• If finalization still fails, the established explicit stalled/max-steps outcome is preserved — the loop never claims success or retries unbounded.

Fixed on top:

• The finalization catch in agent-loop.ts preserved max_steps for ANY error — including an AbortError from the user pressing Esc mid-finalization, which would have relabelled a user cancellation as a stall. The failure is now classified before the guard (same CancelledError/LlmFailure/classifyFailure ladder the loop already uses), so a user abort keeps its cancelled reason and session status. The preserved max-steps path also logs a structured warning instead of breaking silently.

Tests (src/agent/agent-loop.test.ts)

• reserves the final step for a terminal reply (from fix(agent): reserve final step for response #269)
• a user abort during the reserved final inference still ends the turn as cancelled, never stalled/max_steps — fails without the catch fix
• a model that insists on a non-terminal tool at the budget edge gets exactly one repair round-trip, executes nothing, and keeps the explicit stalled outcome (3 LLM calls, 1 tool run, 1 parse_retry)
• maxSteps=1 now deliberately means the single step is terminal: the tool call is rejected before execution and the repair pass must produce the final reply — asserting the behavior change is intentional

Validation

• npm run lint (tsc --noEmit) — clean
• npx vitest run src/agent/ — 13 files, 204/204 passed
• Verified the cancellation regression fails against fix(agent): reserve final step for response #269's code without the catch fix (reason came back non-cancelled), passes with it

Fixes #107