feat: full-corpus reasoning mining — per-model targets + live progress (2.13.0)

[acidkill]

Summary

Run 008 — reworks the reasoning-mining pipeline (introduced in run 007 / #82) so it actually covers the full transcript corpus and gives an operator real control + visibility over distillation, instead of a fixed global cap and a silent, blind background job.

• Deep transcript discovery — recursive rglob discovery finds nested session directories and Task-tool subagent transcripts (projects/<project>/<session>/subagents/agent-*.jsonl), not just the old one-level projects/*/*.jsonl glob. Subagent transcripts attribute to their real top-level project, not a literal "subagents" pseudo-project.
• No per-scan cap — max_traces_per_scan is gone; a full-corpus scan is safe because ingest now inserts per-file (bounded memory) instead of holding every trace in one list. max_trace_chars content-safety limit raised 20k → 100k.
• Backfill = a real full re-scan — --backfill bypasses the incremental size+mtime/line skip and re-reads every file from the top (trace-hash dedup keeps it idempotent), instead of only widening the lookback window. Scan state is never deleted, so a later normal scan stays cheap.
• Per-model pattern targets replace the global run cap — each model gets a distillation target (0–100, dashboard slider or pattern_targets in config/API). A model at target 0 is only detected, never distilled — you set a target, then mining drains that model's backlog to budget.
• Live mining progress — GET /status, the dashboard, the CLI, and docker logs all show phase (scanning → ingesting → distilling → done), files scanned/total, traces found/ingested, and per-model distillation progress.
• opus-4-8 is mined — run 007 wrongly assumed it produced signature-only thinking; the real corpus has ~933 opus-4-8 traces averaging ~1166 chars of genuine reasoning. The denylist mechanism is kept (currently empty) for a future thinking-less model.
• Release 2.13.0 — version bumped across all 8 sources (pyproject.toml, __init__.py, 3 npm manifests + lockfiles), matching the sync convention from fix(ci): sync npm package versions to release tag + unmask publish failures (2.12.1) #86.

Commits (11)

1. feat(reasoning): deep transcript discovery via rglob (nested sessions + subagents)
2. fix(reasoning): remove per-scan trace cap + bump char-safety limit to 100k
3. feat(reasoning): backfill = full re-scan bypass (not state deletion)
4. feat(reasoning): per-file ingest with asyncio.to_thread + progress callback
5. feat(reasoning): per-model pattern targets replace global run cap
6. feat(reasoning): live mining progress in status, logs, CLI, MCP
7. feat(reasoning): mine opus-4-8 (empty the thinking denylist)
8. feat(dashboard): per-model target sliders + live mining progress
9. docs: changelog for run-008 reasoning mining
10. fix(docker): mount host ~/.claude/projects into dashboard for reasoning mining (carried in from fix/reasoning-mining-transcript-mount)
11. chore: release 2.13.0

Test plan

• uv run make verify GREEN throughout (final: 6555 passed, 96 skipped, 1 xfailed, 69.75% coverage, ruff lint/format/security clean, mypy clean)
• BDD acceptance stories authored: qa/run-008/user-stories.feature (local — qa/ is intentionally gitignored per an earlier project decision)
• API tester (real FastAPI app + real SQLite backend + synthetic ~/.claude fixture incl. a nested subagent transcript for opus-4-8): 15/15 PASS — status progress fields, config clamp/422/round-trip, backfill detects all 5 models with total_patterns unchanged at target 0, opus-4-8 mined via the subagent file, raising fable=30/opus=20 distills only those two models, a second normal mine ingests 0 new traces. Evidence: qa/run-008/api/{report.md,run.log} (local).
• Browser tester (real dashboard driven live): 4 range sliders render (0–100 step 10) one per detected model; moving a slider fires a real PUT /config with the exact pattern_targets body; Run mining fires a real POST /mine; post-mine the targeted model shows new pattern/trace counts while every untargeted (target-0) model shows the "set a target and run mining" hint. Evidence: qa/run-008/screenshots/notes.txt (local; PNG capture was blocked by a sandbox renderer issue — DOM + network evidence captured instead).
• Independent review + adversarial ("Ralph") pass, done directly and transparently (not via an opaque background agent): grep confirms no leftover max_traces_per_scan/max_patterns_per_run/TODO/placeholder in shipped code; progress() is never invoked off the event-loop thread; every MiningJobState field has a real write site; background consolidation never receives backfill=True or drain=True by construction. 4 live reproductions (throwaway scripts, actually run): two different models both distill when both targeted (no silent single-model skip); a synthetic subagent transcript attributes to its real project; a legacy config dict with both removed keys loads clean; the drain loop terminates on a no-progress batch.
• Environment verified clean post-testing: smem brain list on the real production SurrealDB shows only default — no orphaned test brains from this session's testing.

Known follow-up (NEEDS-HUMAN, out of scope for this PR)

The real ~3300-file backfill + docker logs INFO-milestone verification on the live docker stack with the actual mounted ~/.claude/projects corpus. This PR's testing used small synthetic fixtures to prove the logic exhaustively; the production-scale run needs the real environment.