fix(sl-hunting): SLH-021 + coverage floor 74% (+ 1 Oct comparison)

[DoRmAmMu1997]

Summary

Three parts:

• SLH-021, a bookkeeping-only code fix. It never places, refuses or sizes an order.
• The 1 Oct trade comparison, doc only. There's no knowledge change, because the one new idea can't be measured.
• Coverage floor raised from 73% to 74% (operator's request).

SLH-021: a trade that opens and closes inside its own decision call was never journaled

The bug. A trade's journal row is written when its decision call finishes (the "harvest"). When the call places an entry and the per-poll stop closes it before that:

• after_exit has no open row to close.
• The stop has already marked the call stale.
• At harvest, SLH-012 sees a stale call and a flat position (it checks self.pos.active), so it throws the call away.

So the trade is missing from both the journal and the decisions log. The runner log shows this four times, each an AI_STOP 3–4 seconds after the fill:

• 10 Aug 09:50:59
• 4 Sep 09:15:48
• 15 Sep 09:15:55
• 30 Sep 09:22:49

Three of those came after SLH-012. Every one of them is a quick loss, so the journal has been making the book look better than it is.

The fix.

• after_exit: when no journal row is open but the executor still holds an unconsumed last_entry_order, the trade must come from the call still in flight, which is the only thing that can place an order. It saves ("parks") what the journal row will need:

  • the closed position;
  • the entry order;
  • the exit details;
  • SLH-020's starting P&L, as the basis for this trade's P&L;
  • the mirror leg's price quality at entry, read before the basket close resets it;
  • whether the mirror is still open.

  It then clears the entry order so it can't leak into a later row.

• Harvest (_consume_agent_decision):

  • It takes the parked trade before any early return, so the trade can never attach to a later call.
  • A stale call with a parked trade is recorded with a WARNING instead of being thrown away, and its decision goes to the decisions log.
  • It opens and closes the row using the parked details plus the call's own decision and market data.
  • If the NIFTY leg was cut alone and the mirror is still open, the close waits for the mirror, as the ordinary path already does. A second entry in the same call can never overwrite that waiting row.

• Known limit: a parked trade whose call is never harvested, because the worker stopped first, is still lost.

1 Oct comparison (doc only)

Both books sold per the note, and both were stopped by BankNIFTY refusing to fall.

• IH: bought puts on all three indices at 09:18, then cut at his loss limit at 10:01. His screen at that point: BankNIFTY +0.62% (54,974.40, just under the note's 55,000), NIFTY −0.12%, SENSEX +0.05%. His loss amount wasn't shown.
• The agent: three shorts, −369.00 in total.
  • Trade 1 booked +3.75 within a minute because BankNIFTY turned up, which was the right read.
  • Trades 2 and 3 then sold into BankNIFTY's round-number rejections, counting them as confirmation. IH says those rejections are what drew in the sellers the breakout went on to hunt.

Why nothing was encoded:

• The mechanism is already written down in R:R-BAIT AT ROUND-NUMBER REJECTIONS, THIRD-INDEX LAG and v4i.
• The only new part — that a rejection in the index refusing your direction is bait there, not confirmation for you — can't be measured, because there's no BankNIFTY price history after 2023.
• The closest thing I could measure was entries taken against the cross-index read: 55 trades, +17,876.75, 51% winners. The two halves split −4,675.75 / +22,552.50, so it doesn't support a rule either way.

Coverage floor: 73% → 74%

• Why it's safe: CI measured 74.6% on main (2c3f4c7, after fix(cpr-ai): prompt v3 - drop the veto examples the backtest refutes #194) and 74.7% on this PR, the same on Python 3.12 and 3.13. That leaves about 0.6 points of margin, roughly 175 statements and branches. That's the same deliberate pressure the 73% floor applied at 0.8 points.
• Where it changed: pyproject's fail_under (with a dated note in its history comment), the repository-policy test, the threshold script's docstring, CLAUDE.md, AGENTS.md, README and the testing-and-ci design doc.

The CI mypy failure (fixed)

• CI runs mypy twice: once on the package files, and once on nifty_multi_strategy_master.py by itself. Locally I had only run the first, which is how this slipped through.
• In _park_close_inside_pass, executor comes from getattr(..., None), so mypy couldn't prove it was non-None in the finally: block.
• The fix is an early return when there's no executor. Behaviour is unchanged, and all ten SLH-021 mutation tests were re-run and still caught.

Test plan

• SLH-021: five tests on the SLH-012 stub, plus four on the real worker. One replays the 30 Sep sequence (enter → invalidate → AI_STOP → stale harvest). It checks that the row carries the call's own setup and reasoning and this trade's basket P&L; 250 of earlier P&L is seeded so the day's total can't pass for it.
• Ten deliberate code mutations were all caught, and a control rewording passes.
• Run during market hours, kept to the touched suites: master suite 648 OK; SL Hunting 283 passed; ruff clean.
• After the two fixes: both mypy runs pass, exactly as CI runs them (python -m mypy on 82 files, then python -m mypy nifty_multi_strategy_master.py). Repository-policy tests: 17 passed. Pre-commit (all files) passes.
• Floor negative test: with pyproject set back to 73.0, the coverage-policy test fails.
• The full local gates (pytest, mypy, compileall, bandit, pre-commit) will run after the 15:15 square-off. Coverage is left to CI, per CLAUDE.md.

🤖 Generated with Claude Code