fix(cpr-ai): prompt v3 - drop the veto examples the backtest refutes

[DoRmAmMu1997]

Why

Prompt v2 gave Codex four example red flags for vetoing a trend-day candidate, but the backtest takes every candidate, so none of the four had ever been tested. I scored each on the 264 five-year candidates (SELL-opposite premium points after costs; baseline PF 1.69):

v2 veto example (as measured)	Trades	Result
Whipsaw: 4+ closes switching side of the running VWAP before the candidate	96	PF 1.71 vs 1.67 for the rest. Vetoing would give up 908 of 2,542 points
Steep push into R2/S2 or a prior-day extreme	6	Five won, about +103 points each: the best group
One climactic bar plus a rejection wick	7	PF 0.81 (−34 points)
Bullish, confluence exactly 2, expiry day	11	PF 1.01 (+4 points)

The size of the candidate bar points the same way. Candidates whose own bar spanned at least 0.25 × ATR5 had PF 2.98 (36 trades). Candidates whose bar alone lifted the range past 1.0 × ATR5 had PF 1.86, against 1.44 for the rest.

Today's paper entry (the first under v2, a single 60-point bar) is the kind of candidate the old climactic-bar example could have led Codex to veto.

What changes

• cpr-trend-day-rider-v3:
  • drops the whipsaw and R2/S2 veto examples;
  • keeps the two rare, roughly break-even ones (the climactic one now says that the wick is the warning, not the size of the bar);
  • tells Codex explicitly not to veto for a big candidate bar, a choppy morning or a push into R2/S2 / a prior-day extreme;
  • quotes the evidence.
• ADR-0019: a dated update (2026-10-01) with:
  • the measurement definitions and results;
  • the 29/30 Sep no-candidate sessions (range 0.90 / 0.75 × ATR5 by 13:30);
  • the 1 Oct entry.
• README and LLD: version references.

Unchanged: the host gate, the VWAP stop, the SELL expression, Codex's veto-only role, and the premise-exit guidance from v2.

Verification

• New test, written first: test_prompt_veto_examples_are_only_the_ones_the_backtest_does_not_refute. It failed against v2 and passes against v3.
• pytest "Tests/Signal Generators/CPR AI Agent": 132 passed.
• Static checks: ruff, mypy (82 files), compileall, and pre-commit on the changed files all pass.
• Not run locally: the full master and other suites. They were skipped because the trading laptop is running a live paper session; CI runs them. This change is prompt text only, and nothing outside the CPR AI suite pins it.

🤖 Generated with Claude Code