baseline: byte-for-byte backtest reference at da61cff + compare tooling

Pin reference outputs of backtest/gpu_backtest.py (XAU/USD, window
2025-08-04..2026-08-04, 4 primary TFs x spread 0.30/0.65 = 8 runs) under
baseline/da61cff/: trade books, stdout summaries, derived equity curves
and metrics, plus a manifest with exact argv, input-parquet and output
hashes, backend, and tool versions.

scripts/capture_baseline.py captures (cache-warm only, aborts rather
than fetch live); scripts/compare_baseline.py re-runs the identical
commands and byte-diffs every output, PASS/FAIL per file. Verified
32/32 files PASS on full re-run at this commit; m5_s065 reproduces the
pinned HEAD reference exactly (555 trades, +98.64R, PF 1.483). No
signal, execution, or scanner logic touched; 359 tests green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J3mZCQRHnJKt7qXQFG3RhB

baseline/da61cff/README.md

@@ -0,0 +1,18 @@
+# Backtest baseline — commit da61cff
+
+Reference outputs of `backtest/gpu_backtest.py` for XAU/USD, window
+2025-08-04 → 2026-08-04, 4 primary TFs × spreads 0.30/0.65.
+Captured by `scripts/capture_baseline.py`; verify with
+`venv/bin/python scripts/compare_baseline.py` (re-runs every command and
+byte-diffs the outputs, PASS/FAIL per file).
+
+Per run dir: `trades.csv` (full book), `summary.txt` (backtester stdout),
+`equity.csv` + `metrics.json` (derived from the book). Those four are
+byte-compared. `run.log` is the backtester's stderr — timestamped, kept
+for provenance only, never compared.
+
+`manifest.json` records the exact argv per run, sha256 of every input
+parquet and output file, backend, tool versions, and the non-secret
+config inputs that alter results. Input data must come from
+`backtest/cache/` — if those parquets change, comparisons are no longer
+attributable to code changes and the compare script aborts.