Stall detection now fires while agent is busy

[bkrabach]

Summary

This branch brings three related improvements to the goal loop's stall detection and observability:

1. Goal progress payload hygiene — The orchestrator now emits the goal condition on every orchestrator:goal_progress event in its fully-expanded form (after file references are resolved), so anyone reading the event sees exactly what the evaluator judged. A schema-version field was added so consumers can distinguish an absent key from a null one. The run summary is now stored in full and truncated only at display time.

2. Evaluator input hygiene — Evaluator input was previously unbounded; a large tool result inside the 40-message window shipped in full every turn. Input is now bounded, tool-call arguments are included (so the evaluator sees what was asked), and the system prompt discloses to the evaluator what it cannot see.

3. Stall detection now fires while agent is busy — Previously a stall was only detectable on a continuation turn that made zero tool calls. This missed the dominant real-world failure mode: the three worst stalls in the corpus (31, 24, and 54 turns) made tool calls every turn while the goal had become unsatisfiable, so the detector never fired.

   A second trigger, independent of tool activity, now runs alongside the original. It uses a cheap deterministic pre-filter over recent evaluator reasons every turn (no model call) and only consults the stall judge when the pre-filter trips, preserving the invariant that a mechanical condition alone never trips a stall.

   The stall judge was reframed from binary yes/no into a verdict taxonomy: resolvable, time-locked, structure-locked, or history-locked. The last three all mean stop now and name the dead end, which users need to rewrite the goal. The judge's prompts were split by trigger, since the original prompt hard-codes the assertion that the assistant took no tool actions.

   The CLI's stalled-state wording now derives from that verdict instead of from a distinct-blocker count. The old counter compared strings with only whitespace and case normalization, so an evaluator that rephrased a blocker each turn produced a distinct signature per turn, yielding a false "flailing" verdict.

   The evaluator reason list is now capped before being passed to the summary model; it previously grew unbounded and a 54-turn run shipped all 54 reasons.

Verification

• Tests: 74 goal tests passed (full app-cli suite has 14 pre-existing failures in test_session_spawner_subprocess, test_cleanup_observability, test_always_render_final_response, test_handler_methods, test_session_lifecycle_events — confirmed present on main before any of this work)
• Real-world measurement: Replayed against the three recorded reason chains, new pre-filter first consults judge at turn 14 (stalled at 31), turn 12 (stalled at 24), and turn 18 (stalled at 54). On achieved organic runs from the same sessions: zero false trips.

Files Changed

• amplifier_app_cli/goal_progress_hook.py — stall detection reframe, verdict taxonomy, input bounding, CLI wording
• tests/test_goal_progress_hook.py — coverage for verdict detection, stall judge confirmation, reason capping