feat(shadow-account): add as-of price-context features (entry RSI, prior return)

[Robin1987China]

Summary

Shadow Account's _compute_features() only learned from journal columns
(holding_days, pnl_pct, entry_hour, entry_weekday, market), so the
extracted rules were purely behavioral with no market context. This adds two
price-context features per profitable roundtrip, read as-of the buy date:

• entry_rsi14 — 14-period Wilder RSI at entry
• prior_5d_return — trailing 5-bar close-to-close return ending at buy_dt

This is the first step you green-lit in #295. I followed the three scoping notes
from that thread:

1. Reused existing pieces. RSI mirrors the causal compute_rsi shape in
   agent/src/skills/technical-basic/example_signal_engine.py:13 (re-implemented
   inline because that hyphenated skills dir isn't importable). Prices go through
   the loader registry (resolve_loader + fetch), the same path
   backtest/runner.py uses.
2. As-of buy_dt, no look-ahead. Both features read only bars dated
   <= buy_dt. The fetch window ends at buy_dt, so the (buy_dt, sell_dt] exit
   window is never touched — look-ahead is structurally impossible, not just
   guarded after the fact.
3. Kept v1 degradation + offline determinism. Any failure (unmapped market,
   no source, empty result, insufficient history) drops the feature to NaN and the
   pipeline still runs. All tests use mocked loaders — no network.

A price feature only joins the KMeans feature set when it's present for
>= min_support roundtrips; otherwise clustering stays journal-only, exactly as
before. Partial NaNs are median-imputed (that affects grouping only — rule bounds
never read price features).

Why

Shadow Account extracts rules from journal columns alone, so extracted rules
are purely behavioral with no market context. Adding price-context features
lets clustering see whether profitable roundtrips share common entry conditions
(overbought RSI, recent momentum), making extracted rules actionable rather
than just descriptive.

Changes

• agent/src/shadow_account/extractor.py: _compute_rsi, _fetch_price_history,
  _as_of_index, _price_features_as_of, _attach_price_features (batches one
  fetch per symbol); _auto_cluster takes a numeric_features arg instead of the
  module constant; _promoted_numeric_features gates promotion.
• agent/tests/test_shadow_account.py: 15 new tests + an autouse offline fixture
  so the suite stays network-free.

Test Plan

• ruff check — clean
• pytest agent/tests/test_shadow_account.py — 43 passed
• full suite (excl e2e) — 4385 passed, 1 skipped
• RSI causality, look-ahead guard (bars after buy_dt don't change values),
  tz-aware buy_dt, all three degradation branches, batch-once-per-symbol,
  median-impute / all-NaN-drop in clustering

Checklist

• No changes to protected areas without prior discussion ([Feature] Trade Journal -> strategy extraction -> backtest bridge (Phase 4c) - planning question #295 green-lit)
• No hardcoded values
• Code follows CONTRIBUTING.md guidelines
• Documentation updated (N/A — no user-facing change)

Refs #295

[warren618]

Thanks @Robin1987China! Merged as b175c25. Nicely PIT-safe — entry_rsi14 / prior_5d_return are read as-of buy_dt through the loader registry, the per-symbol batch fetch still re-slices to each roundtrip's own buy date, and sparse price data cleanly degrades to the journal-only baseline. 🙏