[M4] Rule-based no-progress detector baseline

[askmy-stack]

Summary

• Adds NoProgressDetector in baselines/rules/detectors.py, a rule-based Milestone-4 baseline that flags planning-loop / stagnation failures via two model-independent signals: consecutive repeated actions (same tool name + arguments) and flat goal progress over a configurable window.
• Records goal_progress on each TraceEvent's attributes in TraceCollector.record so detectors can read progress deltas directly from the trace (previously it was only tracked as a parallel list on the collector).
• Adds runtime/features/progress.py with the underlying repeated_action_step / flat_progress_step helpers, following the pattern of runtime/features/mismatch.py.
• Wires the detector into all_detectors() (and thus baselines/rules/__init__.py exports), so it automatically participates in ras experiment run --compare.
• Regenerates experiments/results/tool-semantic-drift-001{.md,-report.json} to show the new detector row (no false alarm on this tool-drift case, as expected — it targets planning loops, not semantic drift).
• Updates the README baseline results table to include the new detector.

Testing

• New tests/unit/test_no_progress_detector.py: repeated-action detection, flat-goal-progress detection, no false alarm on non-consecutive repeats / monotonically increasing progress, registration in all_detectors(), and no false alarms on the existing clean/drifted scripted episodes.
• make ci (ruff lint + format check, mypy, pytest) passes locally — 31/31 tests pass.
• Manually ran ras experiment run --case tool-semantic-drift-001 --compare and confirmed no_progress appears in both the JSON and Markdown reports.

Fixes #22

Test plan

• make ci passes
• ras experiment run --case tool-semantic-drift-001 --compare shows no_progress in detector comparison output
• Unit tests cover repeated-action and flat-progress detection paths plus false-alarm avoidance

Made with Cursor

[askmy-stack]

@askmy-stack Checklist for review:

• Detector implemented (NoProgressDetector in baselines/rules/detectors.py) covering both repeated-action and flat-goal-progress signals
• Unit tests added (tests/unit/test_no_progress_detector.py) — repeated actions, flat progress, no false alarms, registration check
• Wired into all_detectors() and baselines/rules/__init__.py exports
• Appears in ras experiment run --compare output (JSON + Markdown reports regenerated)
• make ci passes locally (lint, typecheck, tests)

Note: --add-reviewer askmy-stack had no effect since askmy-stack is also the PR author — GitHub doesn't allow self-review requests. Assignee was set successfully.