P4-1: persistent Tushare market-data cache (daily + adj_factor) + run-log stats

[StackOverFlow11]

Summary

P4-1: a persistent endpoint-level RAW cache below the feeds so real runs stop refetching full daily bars + adj_factor every time. Disabled by default (backward compatible — an existing config runs exactly as before). The cache stores RAW rows only (unadjusted OHLCV/amount, raw adj_factor); front_adjust still runs in memory downstream, unchanged. PanelStore stays a per-run artifact, NOT the cache source of truth. Market bars only — index_weight/daily_basic/fina_indicator/... are still fetched live (cached in P4-2/P4-3).

• config data.cache (CacheCfg): enabled=False, root_dir (default artifacts/cache/tushare/v1, gitignored), refresh_recent_days=14, force_refresh=[]. Validated; every existing config still loads.
• data/cache/: intervals (closed-day interval algebra — coverage is by calendar range, not row presence, so the planner distinguishes "not fetched" from "source has no row"); parquet_store (per-(endpoint,symbol) parquet under a symbol_prefix shard; atomic upsert-by-(date,symbol), latest wins; best-effort per-symbol lock); coverage (append-only ledger — endpoint/key_type/key/start_date/end_date/fields_hash/fetched_at/row_count/status ok|empty|failed/schema_version/source_version; only ok/empty count as coverage; no token or secret ever written); tushare_cache (read-through: fetch only gaps, upsert, record coverage incl. empty returns; refresh_recent_days refetches a recent tail; force_refresh re-pulls an endpoint; per-run fetch-count stats).
• TushareFeed.get_bars: cache=None → unchanged direct path; cache set → read-through then the SAME join + select so the panel is byte-identical before front_adjust(). Per-symbol retry/throttle stay in the feed via _call closures. pipeline._build_market_cache wires it only when data.cache.enabled.
• Review follow-up (visible hit rate): TushareFeed.cache_stats() + pipeline._log_cache_stats emit one run-log line — data cache: market_daily_gap_fetches=<N> adj_factor_gap_fetches=<M> (cold nonzero; warm historical rerun 0/0).

Real smoke (phase2 baseline)

• Non-cached reference vs cached cold vs cached warm: report metrics IDENTICAL across all three (IC 0.0083 / annual -10.19% / maxDD -16.52% / vol 16.59% / sharpe -0.5703 / turnover 1.0818 / cost 1.19%) — the qfq-equivalence unit test proves cached==direct at the data layer.
• Warm run makes ZERO market_daily / adj_factor calls: coverage ledger 68+68 unchanged between cold and warm; a dedicated warm rerun's run_phase2_baseline.log contains data cache: market_daily_gap_fetches=0 adj_factor_gap_fetches=0.
• wall: ref 998s / cold 1025s / warm 734s (warm saves the market-bar fetches; other endpoints still fetch live). Cache files + ledger carry no token / secret-file content.

Test plan

• pytest -p no:cacheprovider — 411 passed (+28: cache config validation incl. all-configs parametrized; market cache full miss/hit, partial gap, empty coverage, duplicate upsert one-key, qfq equivalence vs direct fetch, no-secret; feed cache_stats() cold {2,2}→warm {0,0}; pipeline logs the exact line)
• ruff check . — clean
• validate-config on all 12 configs — OK
• run-phase0 demo regression — ic 0.9600 / annual 0.8408 unchanged
• git diff --check + conflict marker scan — clean
• AGENTS.md == CLAUDE.md (verbatim copy convention)
• secret scan on the merge diff — 0 token, no artifacts/parquet committed
• real smoke run (3-run metrics identical + warm 0/0 in the run log)

Invariants preserved: front_adjust formula/timing unchanged (qfq-equivalence test + 3-run real evidence); factor/alpha/portfolio/execution/OOS/report math untouched; artifacts/data/{output_name}.parquet not treated as the cache SoT.