feat(rl): fractal-zigzag entry plug-in on 3-minute Parquet

[lgbarn]

Summary

• New RLStrategy plug-in futures_foundation/rl/fractal_zigzag.py: detect_entries driven by detect_fractal_zigzag_pivots (the trigger-scan causal winner) with live_edge=True newest-bar confirms, so entries are truncation-invariant at the data edge.
• Every candidate carries direction, detection-time entry reference price (close of the confirm bar; actual fill = next bar's open, enforced centrally by the pipeline), sl_distance = 1x Wilder ATR(20) at the signal bar, and the absolute stop_price.
• compute_obs_features reuses the causal pivot-structure block from futures_foundation.pivots (htf_dir excluded — its ATR floor is a full-series median, not truncation-invariant, so not obs-safe).
• load_3min_parquet + SYMBOLS for the six-instrument 3-min Parquet data; registered in the RL registry as "fractal_zigzag".

Closes

Closes #7

Acceptance criteria

• Entries are truncation-invariant: running detection on a truncated series yields identical entries up to the truncation point — test_entries_truncation_invariant (generic check_causal harness + exact prefix==full frame equality)
• Every candidate entry has a 1x ATR initial stop attached at detection time — test_every_entry_has_1x_atr_stop
• All six symbols load from Parquet and produce plausible entry counts over the full 5-year span — test_six_symbols_load_and_produce_plausible_counts (data-gated); counts reported below
• Feature computation is causal — test_obs_features_truncation_invariant (byte-identical rows under truncation)

Entry counts (full span, k=2, min_leg_atr=1.25, ATR(20))

symbol	bars	span	entries	long/short	bars per entry
NQ	602,765	2021-04-28 .. 2026-06-04	80,759	40,380 / 40,379	7.5
ES	602,766	2021-04-28 .. 2026-06-04	62,849	31,425 / 31,424	9.6
RTY	601,700	2021-04-28 .. 2026-06-04	73,327	36,664 / 36,663	8.2
YM	602,656	2021-04-28 .. 2026-06-04	74,138	37,069 / 37,069	8.1
GC	603,710	2021-04-28 .. 2026-06-04	69,552	34,776 / 34,776	8.7
SI	602,571	2021-04-28 .. 2026-06-04	59,828	29,914 / 29,914	10.1

The near-exact long/short split follows from the detector's strict pivot alternation; the ~1-entry-per-8-10-bars density is the intended mechanical-candidate stream (the PPO chop-veto is the filter, per the RLStrategy design).

Test plan

• uv run pytest tests/test_fractal_zigzag_strategy.py -q — 4 pass (parquet test skips without data)
• FFM_DATA_DIR=<path-to>/data uv run pytest tests/test_fractal_zigzag_strategy.py -q — 4 pass including the six-symbol span/count check
• uv run pytest tests/test_fractal_pivots.py tests/test_rl_pipeline.py -q — adjacent suites unchanged, green

Notes for reviewer

• TDD seam record (issue carries the tdd label): each acceptance criterion was one red→green cycle, in commit order — truncation invariance → 1x ATR stop → causal obs features → Parquet loader/counts.
• live_edge=True is hardwired in detect_entries: it is both the issue's requirement (newest-bar confirms for bar-by-bar consumers) and what makes the prefix==full equality exact at the boundary (a training-mode cf+1 < n guard would drop the newest pivot on the prefix side).
• - [x] items are author claims verified by local test runs; the parquet-gated test skips on CI (data is local-only/gitignored), so the counts table is the CI-visible evidence for AC 3.
• Footprint: two new files only (futures_foundation/rl/fractal_zigzag.py, tests/test_fractal_zigzag_strategy.py); rl/env.py, rl/__init__.py, tests/test_rl_pipeline.py, and topstep.py untouched (sibling-owned this wave).
• The local full-suite run shows 39 pre-existing failures, all ModuleNotFoundError for optional deps (sklearn/torch/optuna/onnxruntime) absent from the local env; CI installs requirements.txt and is the arbiter.

Out of scope

• Wiring this strategy into a walk-forward/PPO run config (that is the pipeline side, issue PRD: PPO combine-manager on fractal-zigzag triggers — Topstep 100K Optuna objective #1 tree).
• HTF direction as an observation feature — would need a causal (expanding/frozen) ATR floor in causal_htf_dir first; noted, not changed.