… produced each optimizer attempt — required for the scorecard screen

…oint groups candidates by stage and picks the best.

Frontend shows summary cards, per-stage ParityBeam scores, and expandable
rows with prompt variant text and all attempt detail

…ut schema

…runner tests

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- New PipelineGraph component (pipeline-graph.tsx): stage nodes via dagre,
  orientation toggle (→/↓) persisted per-pipeline, calls drawer on node click
- ModelPanel floating card highlights stages that share the same LLM, giving
  the 'Obsidian node graph' cross-connection insight without layout conflicts
- Backend: StageInfo now includes trace_count and total_cost_usd (func.count +
  func.sum added to the /dag aggregation query)
- api.ts StageInfo updated to match; all test fixtures updated with new fields
- TESTING.md §3.3d added for Graph tab walkthrough

- ModelGraphNode in fixed right column, dashed edges from each stage
- Click model node highlights/unhighlights connected stages
- Click stage node (with traces) expands call nodes inline below it
- CallGraphNode shows tokens/latency/cost; straight edge from stage bottom
- Records fetched lazily on first expand, cached for re-expand

…ion)

Phase 2: train/test holdout split, holdout_score on Candidate, StageResultOut, migration-detail Holdout column
Phase 3: GET .../export JSON attachment, Download winning config button (fetch+blob)
Phase 4: seam.py evaluate_seam, SeamCheckResult model+migration, _run_seam_regression, seam-results endpoint, Seam checks table
Tests: api 191 passed | core 297 passed 21 skipped | web 153 passed

…ertions API + UI

…rections

Fixed a real bug found in code review: _run_seam_regression hardcoded
parity_threshold=0.95 instead of using the migration's actual configured
value, while the main optimization loop already threads it through
correctly elsewhere in the same file. Now passed as a required keyword
argument from migration.parity_threshold at the one call site.

Corrected docs/TESTING.md's Graph tab section, which described a floating
panel and a slide-in drawer that don't exist - the real implementation
renders everything as inline React Flow nodes.

Rewrote START_HERE.md's badly stale "Not built yet" list - items 1-5 were
all long done (rubric trigger, model-card info, the full M3 optimizer
loop, M4 runs with live status, M5 config export/holdout/seam) but the
list still described them as pending.

apps/api: 191 passed, unchanged count (behavioral fix, not new coverage -
see DEV_TRACKER.md's noted follow-up on the orchestration-glue test gap
this bug slipped through).