fix(fetch_fred): warm S&P 500 covariates and skip delisted gold series

[ethancjackson]

Summary

Unified fix for FRED-related items. fetch_fred.py only warmed the food-price series, so the S&P 500 implementation's FRED covariates were never cached — even though fetch_sp500_market.py explicitly defers to it for them.

• Warm both experiments' covariates. Import FRED_PREFETCH_REGISTRY from sp500_forecasting.data (whose own comment already declared it was "for scripts/fetch_fred.py registration" — it was just never consumed) and fetch the union of the food-price and S&P 500 series. FRED ids are unique across the two sets, so no de-duplication needed.
• Per-series frequency. Catalogue entries now carry a frequency field, so daily S&P 500 series (DGS10, DGS2, DFF, DCOILWTICO, DTWEXBGS) register as D instead of the old hardcoded MS. The S&P 500 covariate builders handle alignment themselves; this script only warms the raw parquet cache.
• Delisted gold series. Both London gold fixing series (GOLDAMGBD228NLBM, GOLDPMGBD228NLBM) are permanently discontinued by FRED (HTTP 400 "series does not exist") with no equivalent daily USD replacement. Marked them known-unavailable via KNOWN_UNAVAILABLE_FRED_IDS so a clean run reports 13 registered (0 failed, 2 skipped) instead of two spurious [failed] lines. Documented the discontinuation in sp500's registry where the ids live. No modeling behavior change — gold stays in the covariate list and still degrades to absent at runtime via the existing first-available fallback (strict_covariates=False).
• Regenerated the concierge-context mirrors for the two changed source files.

Verification

• uv run python scripts/fetch_fred.py → Registered 13 series (0 failed, 2 skipped).
• ruff check passes on both source files.
• pytest implementations/tests/sp500_forecasting/test_data.py aieng-forecasting/tests/aieng/forecasting/data/test_fred_adapter.py → 7 passed.

Out of scope (deliberately)

• Regenerating the concierge context surfaced ~25 unrelated stale mirrors + a volatile catalog-summary.yaml from earlier merged PRs. That mirror set isn't regenerated per-PR, so this PR keeps only the two mirrors it actually changed — a full refresh belongs in a separate maintenance PR.
• No daily FRED gold replacement exists (GVZCLS is volatility, not price; IR14270/ID7108 are monthly indices that don't fit the daily log-return builder). A real substitute (e.g. GLD via yfinance) would be a modeling change, not a quick win.

🤖 Generated with Claude Code