PR4 (data, DRAFT): delete slim writer + load_slim_cache API + fallbacks [merge-gated 5/23]

[cipher813]

Terminal PR of the Wave-4 predictor/price_cache_slim/ deletion arc (data side).

Status (revised 2026-05-25)

Original framing (when opened 2026-05-19): merge-gated on a 5/23 Saturday-SF WAVE4_PARITY_METRIC passed: true read across the breadth + compute + exit_timing streams.

Why that gate is the wrong question. The remaining slim ↔ ArcticDB divergence (5/24 SSM log: max_abs_value_delta: 5.36 on EQIX Close 2026-04-29, 12,105 cells over 1e-6 epsilon) is by-design adjustment-policy mismatch, not a defect signal: ArcticDB universe writes auto-adjusted close (yfinance default), slim writes raw close. Both are correct representations; neither is "the" answer. EQIX is a REIT with $5+ quarterly dividends — over the comparison window the cumulative dividend adjustment IS the delta. Even after data #305 relaxed epsilon to 1e-2, passed is gated on n_cells_over_epsilon == 0 (alpha_engine_lib/reconcile.py:62-68); since dividend-scale deltas remain > $0.01, epsilon alone cannot flip the gate.

The actual safety case is stronger than the gate could ever provide:

1. ArcticDB cutover landed 2026-04-14; consumer migrations have been ArcticDB-primary in production since their respective merges (data PR1a: macro breadth → ArcticDB (slim fallback + parity-observed) #267 macro-breadth, data PR1b: feature compute price+macro → ArcticDB (universe+macro libs, slim fallback, parity) #268 feature-compute, backtester feat(news_sources): adapter implementations + aggregator (Wave 1 PR β; requires lib v0.15.0) #226 exit_timing, dashboard fix(ci): grant GHA role full resource CRUD for CF stack updates #88 health-check). 6 weeks of weekly Saturday SF + daily inference + weekly backtest cycles, no correctness alerts.
2. The institutional question — "should slim go away?" — was decided affirmatively when the ArcticDB cutover landed. The parity gate's job was to verify ArcticDB had equivalent data BEFORE we deleted the writer; the consumer-side soak provides that evidence with more rigor than the cell-equality check.
3. Merge is reversible. Revert PR restores the writer + fallbacks. The irreversible step is aws s3 rm --recursive per the runbook, and that's separately gated by a byte-equal S3 backup (Wave-5 precedent).
4. Slim is a 2y slice derived from predictor/price_cache/, not source data — regenerable from full cache or ArcticDB at any time.

Code removed

• collectors/slim_cache.py — the 2y-slice writer — deleted.
• weekly_collector.py — slim import + # 3. Slim cache collect block + --only slim choice + docstrings.
• store/parquet_loader.py — load_slim_cache + SLIM_CACHE_PREFIX + now-unused ThreadPoolExecutor. load_parquet_from_s3 kept — it's re-exported via features.compute._load_parquet_from_s3 and consumed by builders/backfill.py.
• collectors/macro._load_breadth_prices → ArcticDB-only. No-null breadth contract preserved (None → key omitted) on ArcticDB failure.
• features/compute._load_price_source → ArcticDB-only (universe + macro libs). No-data contract preserved.

Tests

• test_macro_breadth / test_compute_price_source reworked to the ArcticDB-only terminal state (slim-fallback/parity tests retired).
• test_wave4_slim_arctic_parity.py repurposed from migration scaffolding into a permanent regression guard: the functional slim surface (writer/loader def, call, or import) must never return.
• Full data suite passing post-rebase.

Runbook (not executed by this PR)

WAVE4_SLIM_DELETION_RUNBOOK.md — the gated, manual S3 prefix deletion: pre-deletion byte-equal backup (Wave-5 precedent) → aws s3 rm → verify-empty.

Companion PR

alpha-engine-backtester #227 — removes the slim leg from exit_timing._load_price_cache + the verified-dead spot_backtest.sh slim sync. Merge after data #269 + Lambda canary green.

🤖 Generated with Claude Code