PR1b: feature compute price+macro → ArcticDB (universe+macro libs, slim fallback, parity)

[cipher813]

The riskiest consumer migration of the Wave-4 arc (predictor/price_cache_slim/ deletion). Consumes lib v0.20.0 (PR0a-2 lib #51, merged: load_macro_series + shared read core).

Why this one is risky

features/compute._load_prices_and_macro feeds the entire feature-compute pipeline and _extract_macro. The slim cache historically carried equities + SPY + index/macro series (VIX/VIX3M/TNX/IRX/GLD/USO) + all XL* sector ETFs in one flat dict. Those tenants are split across two ArcticDB libs — a naive load_universe_ohlcv swap would silently drop every macro/ETF series → _extract_macro near-empty → SPY/VIX/VIX3M-derived features degrade across the whole universe. (This is the audit finding that justified the separate PR + the PR0a-2 load_macro_series helper.)

Change — composed read

New _load_price_source():

• load_universe_ohlcv(bucket) → equities + SPY
• load_macro_series(bucket, syms) → _MACRO_SLIM_KEYS ∪ XL* (discovered via open_macro_lib().list_symbols(), startswith("XL"); the heterogeneous non-price features key is excluded by load_macro_series's explicit-symbols contract)
• union = the slim-equivalent dict _extract_macro + the pipeline consume unchanged
• slim retained as whole-set fallback — feature compute cannot run blind; returns None (→ empty, existing no-data contract) only if both sources fail
• SOTA observation: while both exist, every run dual-reads + emits reconcile(...).as_metrics() JSON — grep WAVE4_PARITY_METRIC compute. require_ticker_match=False because slim legitimately carries symbols the universe lib does not — set asymmetry is logged in the metric fields for visibility while passed reflects value fidelity over the overlap. Dual-read + slim removed in PR4.

Safety

• No Dockerfile-extra change (precise audit): features/compute.py is not Lambda-packaged — the Dockerfile doesn't COPY features/; it runs only via weekly_collector on EC2 spot, whose requirements.txt already carries [arcticdb,flow_doctor,rag].
• Pin bump 0.19.0 → 0.20.0 (requirements.txt + Dockerfile, lockstep guard enforced).
• +5 tests (compose / slim-fallback / parity-emit / both-fail / empty-source). Full data suite 1380 passing; Wave-4 anti-drift consumer-set guard still holds (compute.py keeps slim fallback → stays in WAVE4_INVENTORY until PR4).

Sequence

This completes the PR1 (consumer migration) tier on the data side — both data-repo data-read consumers (macro.py PR1a #267 merged; features/compute.py here) now ArcticDB-primary + parity-observed. Next: PR2 (backtester exit_timing.py), PR3 (dashboard health_checker.py retire), PR4 (delete writer+prefix, gated on one clean Saturday-SF zero-diff ParityReport across the WAVE4_PARITY_METRIC streams).

🤖 Generated with Claude Code