feat: active coaching loop — evidence sentences, score caps, trend deltas, milestones

[huccct]

Summary

• Evidence sentences: every dimension scored above 5 must cite a specific prompt behavior; no evidence = score capped at 5.0
• No-evidence cap: deterministic engine cap prevents inflated scores without verifiable justification
• Per-dimension trend deltas: trend entries now show how each dimension changed vs the previous session
• Earlier trend display: trend and focus area shown from session 2 (was 3)
• Milestone detection: engine detects sessions 5, 10, 20, 50, 100 and surfaces them in the report
• Report restructure: flat 8-dimension list replaced with Focus Dimension + folded others + Next Step conversation offer
• Score-triggered rewrite: /rewrite-last now supports being triggered directly from a /score report

Test Plan

• 70 tests passing (12 new tests added)
• Engine smoke test: evidence, milestone, recent_trend with dimension_deltas all present in output
• No regressions on existing test suite