feat(sl-hunting): v5n + SLH-020 - the second seller is the random seller (+ 30 Sep note)

[DoRmAmMu1997]

Summary

Most of this PR comes from the 29 Sep session: IH's live video (6zppfK_n4yg) compared with this book's two trades.

• v5n is a knowledge-only rule.
• SLH-020 (operator-approved addendum) is a small runtime change. It only adds information for the agent and never refuses an order.
• The 30 Sep pre-open note (addendum).
• The 26 Sep weekly (addendum): doc only. Its one new idea failed a data check, so nothing was added to the prompt.

This was also v5m's first live day: the runner started at 09:14:46, after #190 merged, and logged the injected note.

	IH	agent
first leg	sold from 09:20, booked 09:30	09:22 → 09:32, +6,246.75 (the same trade)
second entry	none	10:02 into the range the first leg left behind, cut 10:35 with NIFTY flat, −1,504.50

The agent contradicted itself three minutes apart:

• 10:00:51, holding: "that first leg already played out… consolidating in a tight ~50pt range".
• 10:02:57, entering: "declined continuously… no major retracement".

None of the day's 80 recorded decisions cited v5m.

v5n: THE SECOND SELLER IS THE RANDOM SELLER

• Placement: a sub-bullet of MOVE-EXHAUSTION, right after v5m.
• What it adds: IH explains why v5m's losses happen. On a follow-the-market day, the sellers who arrive late are "random sellers" who cut at the first candle that goes against them, and the next retracement exists to remove them. So a second same-side entry after a winning leg joins exactly that crowd.
• What it doesn't say: IH also said "opportunity comes, in an hour or two, but at the right place", about his own late entry. So the rule is not "never re-enter".

Re-measured (tested, not enforced)

• After a winning exit: 36 re-entries, −10,807.25. After a losing exit: 43 re-entries, +33,714.25.
• Both halves of the after-win set are now negative (−9,312.50 and −1,494.75), but it still isn't robust:
  • the most recent third made +139.50;
  • without its two worst sessions the set would total +1,414.75;
  • the gap to after-loss re-entries is p = 0.056, before correcting for the other splits tried.
• So this stays a default, not an enforced gate.

SLH-020 (addendum): tell the agent about today's last close when it re-enters the same side

The prompt rule didn't hold on its first day, and the numbers don't support an enforced gate, so this is the middle option. It refuses nothing.

• Worker: records each finished trade at the moment the NIFTY + BankNIFTY basket goes flat. Every close path passes through that point: the agent's EXIT, stop/target, max-loss, square-off, and closing a lone mirror leg. The record holds:
  • direction and close time;
  • booked_pnl: the change in realized_pnl since entry, both legs (the same basis as SLH-019);
  • nifty_when_flat: read from the shared store only, so there's never a broker call on the exit path.
    It's recorded even when the post-exit cooldown is turned off, and last_closed_trade_today() only returns today's record.
• Executor: while flat, position_state carries last_closed_trade_today. An accepted entry on the same side as a winning close from earlier today carries same_side_after_winning_exit, and it logs an SLH-020: line so flagged entries can be counted later. The existing SLH-005 cooldown and SLH-016 already-breached-stop refusals are unchanged.
• Descriptions: the tool descriptions and the v5n rule both name the two fields and state that they refuse nothing.

Addendum: pre-open note for 30 SEP 2026

Source: NrPmCbzrzSY, uploaded 29 Sep 21:25 IST. The branches flipped from 29 Sep:

• Flat to a small gap-down → BUY. The buyers that the recovery drew in have already been retraced out, and the market is trying to go up again.
• Gap up → SELL. No sellers' stops sit above to fuel a rise, and others may buy the gap reading it as positive.

index	resistance	support
NIFTY	22900 / 23000	22650 / 22560 (same lines as 29 Sep)
BANKNIFTY	54650 / 55000	54000 / 53790
SENSEX	73130 / 73450	72000 / 71500

Two support pairs were garbled by the transcript, so I read them off the 1080p chart:

• BankNIFTY "5453 790" is 54,007.90 / 53,787.80.
• SENSEX "7271500" is 72,000.23 / 71,510.61.

There's no expiry today.

Addendum: the 26 Sep weekly — one candidate, not encoded (doc only)

Source: dVCXZ1PXVYI, 13:32, read in full. Nearly all of it is already in the knowledge:

• his buyer chart is RECRUITMENT HISTORY;
• "flat, gap-up, even a small gap-down, same plan" is A SMALL GAP DOES NOT RESCUE A SEATED CROWD;
• his 25 Sep loss is already in v5l.

The one new idea is a split between two kinds of seated sellers:

• sellers from one sudden fall would flee an adverse gap-up "in one or two candles", so that branch is not a hunt;
• sellers seated through a multi-day fall that already retraced would stay huntable in every open.

Measured on five years of NIFTY spot 1-minute data, the gap-ups did the opposite:

prior session	gap-ups	median rise left after 09:16	mean net to 10:30	net up
one sudden fall	10	65.65 pts	+6.59	60%
multi-day fall that retraced	21	37.00 pts	−13.96	38%

The result held in five threshold variants. The samples are small and the definitions are rough stand-ins, so it doesn't prove the reverse either. But I wasn't willing to put a rule the data contradicts into a live prompt. The doc records the lecture, the idea and the measurement, plus a transcript workaround this video needed.

Test plan

• test_v5n_the_second_seller_is_the_random_seller pins seven clauses, plus the SLH-020 wiring. Nine deliberate mutations of the rule were all caught, and a control rewording passes. The first mutation run found one phrase the test wasn't checking, so I tightened the test and re-ran the full set.
• SLH-020, on stand-in workers (six tests):
  • the flat snapshot reports the close;
  • a same-side entry after a win still reaches the worker and carries the flag;
  • no flag for the opposite side, a losing close, a break-even close, a non-numeric P&L, or when nothing has closed today;
  • a broken hook changes nothing;
  • a cooldown-refused entry is refused exactly as before;
  • the tool descriptions name the fields.
• SLH-020, on the real worker (three tests):
  • a mechanical AI_TARGET close is recorded with the basket's booked figure. An earlier P&L is already booked that day, so the running total can't pass for this trade's figure. The next same-side entry is flagged (cooldown off);
  • a losing close is recorded but never flagged;
  • a record from another session is never reported.
• Nine deliberate code mutations were all caught, and a control passes.
• Pre-open note: the note test was retargeted to 30 Sep. Seven deliberate mutations of the note were all caught (wrong date, un-flipped branches, each branch's reason, a stale BankNIFTY support, a garbled SENSEX support, yesterday's source id), and a control rewording passes. SL Hunting suite: 283 passed; pre-commit passes.
• Full local gates, run after the 15:15 square-off: master suite 639 OK; market-data-health 28 OK; pytest 1701 passed; ruff, mypy (82 files) and compileall pass; bandit 0 findings; pre-commit passes. Coverage is left to CI, per CLAUDE.md. These gates ran before the two addenda, which change only the note, its test and the doc.

🤖 Generated with Claude Code