feat(rl): 3-seed confirmation + go/no-go verdict report

[lgbarn]

Summary

• Negative result, reported honestly: the sweep winner fails every PRD acceptance gate on the holdout, and the verdict is ITERATE — do not pay for a combine on this evidence.
• Headline numbers (holdout test months 2025-06 -> 2026-06, 13 walk-forward windows): pooled pass rate 0.35% (3/852 attempts) vs the 60% bar; 813 busted-MLL; per-seed 0.7% / 0.0% / 0.4%; and the policy underperforms both no-skill baselines (random-take 11.5%, take-every-signal 5.2%) — the report says so plainly and marks the verdict DISCARD-leaning.
• New futures_foundation/rl/seed_confirm.py: reruns the logged Optuna winner (from study storage via --storage, or explicit --params) on N seeds with only the seed varying over the same 3m/1m walk-forward protocol the sweep used, on holdout test months the sweep never touched; runs both no-skill baselines through the identical evaluate/summarize seam; scores the three PRD gates (>= 60% pooled pass rate, both baselines strictly beaten, every seed >= 50% with zero blowups).
• New self-contained verdict report docs/topstep_100k_verdict.html (inline CSS + inline SVG, zero external references): gate scoreboard, bust breakdown, per-symbol attribution, attempt counts, winning knobs, explicit ITERATE verdict with next levers.
• Tests: tests/test_seed_confirm.py — pure gate/verdict seams with stub runners (no SB3), following the test_optuna_sweep.py pattern.

Closes

Closes #5

Acceptance criteria

• 3-seed retrain runs from the logged winner configuration with only the seed varying
• The report states pass/fail against every acceptance gate from the parent PRD (>= 60% pass, baselines beaten, all seeds >= 50% with zero blows)
• Report is one self-contained HTML file in the repo docs (inline CSS/SVG, no external references)
• The verdict line is explicit: Ship / Iterate / Discard, with the one-paragraph reason

The real run (evidence)

Winner: Optuna trial 19 of study topstep-100k-v1 (40-trial sweep, 2023-01-02 -> 2025-06-01, 26 walk-forward windows, 10k PPO timesteps per window): activate_r=0.554, trail_atr_k=2.518, dll_penalty=1.769, mll_penalty=0.650.

Confirmation command (run 2026-07-12; winner knobs at 3-decimal precision via --params — --storage/--study-name re-derives the full-precision equivalent):

uv run python -m futures_foundation.rl.seed_confirm \
    --data-dir data \
    --params '{"activate_r": 0.554, "trail_atr_k": 2.518, "dll_penalty": 1.769, "mll_penalty": 0.65}' \
    --start 2025-03-01 --end 2026-06-04 \
    --timesteps 10000 --seeds 0 1 2 \
    --out seed_confirm_results.json

Real output (per-run summaries + gate table, from seed_confirm_run.log):

-- policy seed 0 --
windows=13  signals=89814  taken=28369  skipped-while-open=60869
attempts=304  passed=2 (1%)  busted=290 [busted-MLL=290]  timeout=12  median-days-to-pass=3.0
-- policy seed 1 --
windows=13  signals=89814  taken=27872  skipped-while-open=59727
attempts=278  passed=0 (0%)  busted=266 [busted-MLL=266]  timeout=12  median-days-to-pass=n/a
-- policy seed 2 --
windows=13  signals=89814  taken=27205  skipped-while-open=58986
attempts=270  passed=1 (0%)  busted=257 [busted-MLL=257]  timeout=12  median-days-to-pass=6.0
-- baseline random-take --
attempts=87  passed=10 (11%)  busted=64 [busted-MLL=64]  timeout=13  median-days-to-pass=6.0
-- baseline take-every-signal --
attempts=116  passed=6 (5%)  busted=97 [busted-MLL=97]  timeout=13  median-days-to-pass=6.0
== PRD acceptance gates ==
gate 1  pooled pass rate 0.4% (bar 60%, 852 attempts)  FAIL
gate 2  baselines beaten (random-take: no, take-every-signal: no)  FAIL
gate 3  seed stability (seed 0: FAIL, seed 1: FAIL, seed 2: FAIL; 813 busts total)  FAIL
VERDICT: NO-SHIP — Iterate vs Discard is the report's call

Test plan

• uv run pytest tests/test_seed_confirm.py tests/test_optuna_sweep.py tests/test_multi_combine.py -q — 25 passed, 2 skipped locally
• Open docs/topstep_100k_verdict.html directly from a clone — renders fully offline, light and dark
• Smoke the CLI: uv run python -m futures_foundation.rl.seed_confirm --data-dir data --storage sqlite:///optuna_topstep_100k.db --study-name topstep-100k-v1 --symbols NQ ES --start 2025-03-01 --end 2025-07-15 --timesteps 1000 --seeds 0 (minutes)

Notes for reviewer

• This PR reports a negative result on purpose. All 40 sweep trials scored ~0% combine pass rate (best objective -544; every completed trial busted accounts). The 3-seed confirmation exists to make that verdict solid, not to rescue it; the gates fail and the report says ITERATE with concrete next levers (more PPO timesteps per window, wider reward-shaping ranges, entry-quality filter; the study is resumable so round two extends it).
• Existing modules untouched: the runner recombines optuna_sweep / multi_combine / pipeline seams. Only new files are added.
• 3 pre-existing local failures in tests/test_evaluate_overfit.py (pooled-AUC tests) reproduce on the base commit and are environment-only — CI on main is green.
• The Optuna study DB (optuna_topstep_100k.db) and raw run JSONs stay out of the repo (run artifacts, kept locally); the report embeds all numbers.
• Gotcha for anyone rerunning this: the data/*.parquet training inputs are gitignored — a fresh clone or a git worktree has no data/ directory. Restore the parquet files (or point --data-dir at a checkout that has them) before running; the report's Reproduction section says the same.

Out of scope

• Round-two iteration itself (extending the study, longer training, entry filters) — the report's "Next levers" section is the plan of record.
• Fixing the pre-existing test_evaluate_overfit.py local-env failures.