feat(signal_returns): write calibrator-v1 context on score_performance seed + backfill

[cipher813]

Summary

Closes the producer-side root cause behind backtester #176's evaluator P0. Two writers had diverged silently after research migration #12 (2026-05-08): `scoring/performance_tracker.py::record_new_buy_scores` wrote all 5 canonical context columns but had zero production callers; `collectors/signal_returns.py::_seed_score_performance` was the actual production writer (runs weekly in DataPhase1) and only wrote `(symbol, score_date, score, price_on_date)`. Result: every row in `score_performance` post-migration had NULL `quant_score` / `qual_score` / `conviction` / `sector_modifier` / `market_regime`.

Production audit at time of writing:

date	rows	quant_score populated	qual_score populated
2026-04-30	23	0	0
2026-05-01	23	0	0

(`investment_thesis` for the same dates had 13/15 and 50/58 populated, so the upstream LLM pipeline IS producing the values — they were just being dropped at the seed step.)

Changes

• `_seed_score_performance` now reads sub_scores + conviction + sector + market_regime from the same signals.json payload that drives the BUY filter (no second round-trip; single source-of-truth fetch per date).
• `_backfill_score_context` (new) repairs legacy rows whose canonical columns are NULL. UPDATE-WHERE-NULL means re-runs converge to a no-op once every row has at least one source. Wired into `collect()` as Step 1b.
• `_ensure_score_performance_schema` defensively mirrors research migration Add GitHub Actions auto-deploy on push to main #12 (5 calibrator-v1 context columns). Belt-and-suspenders in case DataPhase1 ever fires against a fresh research.db before research's cold-start migrations.
• `_extract_signal_context` (new helper) matches the extraction logic in alpha-engine-research's `scripts/backfill_calibrator_v1_context.py` so producer-side seed and legacy backfill agree on every field's source.

Test plan

• `pytest tests/test_signal_returns_score_context.py -q` → 11 passed (3 extraction cases, 3 seed-INSERT cases, 4 backfill cases, 1 schema-ensure idempotence case)
• `pytest -q` (full suite) → 705 passed, 1 skipped, 0 failures
• Saturday 2026-05-16 SF DataPhase1 emits canonical context on newly-seeded rows
• Same SF run's backfill step closes the 186-row legacy NULL gap in production research.db
• weight_optimizer status=ok end-to-end on that same SF run (no auto-rollback from the column-rename class)

Composes with

• alpha-engine-backtester feat(alarms): per-row + aggregate CloudWatch alarms for transparency substrate #176 (consumer-side coalesce fix, shipped 2026-05-10) — with this producer fix in production, the next backtester PR can retire the S3 round-trip in `weight_optimizer.load_with_subscores` entirely (score_performance becomes the sole authoritative source).

🤖 Generated with Claude Code