feat(sl-hunting): v4h knowledge from the 14 Aug IH live session

[DoRmAmMu1997]

What

v4h of the SL Hunting knowledge, distilled from the 14 Aug Intraday Hunter live
session (H7RXIw5xbZk) — a win held through a full drawdown. Every prior
version learned from a clean win or a clean loss, so the prompt had no worked
example of a correct trade that looked wrong in the middle.

Knowledge-only. No runtime behaviour changes.

The two rules

ENTRY QUALITY AND DIRECTION ARE SEPARATE JUDGEMENTS — a green result does not
retroactively make an entry good, and an immediate adverse move does not make the
side wrong. Without the split, the journal teaches the agent to grade its reads by
their P&L, which is the loop that produces revenge entries after a stopped-out
correct call.

THE LOSS LIMIT IS A PERMISSION TO WAIT — the reason this version exists, and
the most dangerous line in the prompt if read alone: "permission to wait" is one
careless paraphrase from "hold losers". It is bounded on both sides in the prose
and both bounds are asserted in the test — it licenses patience toward the stop,
never past it, and forbids cutting inside it on feel. It names v4e's
DISCIPLINE IS ASYMMETRIC as the rule it refines rather than overrides.

v4g's time rule is extended rather than left to contradict that patience: time
does not merely shrink the payoff, it EXPIRES THE PREMISE. Elapsed time can
retire a premise before price ever reaches the limit.

⚠️ The prompt budget is now the binding constraint

assembled                :  114,021
budget for assembled     :  114,140
under by                 :      119

119 characters. The append-only rhythm this series has used since v3 is finished.

Two consequences, both deliberate and both recorded in the doc:

1. One rule was cut for budget, not merit. A SEATED CROWD IS WARNED BEFORE IT IS HUNTED, SO NO WARNING MEANS NO CROWD was written and tested, then removed to
   fit. It is the sharpest new idea in the session — an early adverse spike is the
   market shaking a seated crowd, so its absence confirms an empty book. It should
   be the first thing restored once there is room.
2. v4i cannot ship anything without pruning first. OPENING_DRIVE and RISK
   have both accumulated rules that later versions restated with better triggers.
   That is a knowledge-quality task deserving its own change, not a rushed trim at
   the end of a version — so it is not done here.

This PR does prune one: v4e's A SHARP FIRST SLIDE BAITS; A SLOW ONE MEANS IT,
superseded by v4f's confirmation rule, which covers the same tell with a
checkable trigger.

Agent vs IH on the same session

Basket realized −3,854.50 across 47 legs. SL Hunting AI: +2,866.50 over 2
trades — best strategy on the board for the second consecutive day, cross-checked
against its own Result summary.

Time	Decision	Reason
10:04	ENTER_LONG	flush_reversal_hammer_confirmed
10:07	EXIT	double_top_rejection_premise_fade
10:20	ENTER_SHORT	double_top_bearish_engulfing_breakdown

Both the agent and IH finished on the same side (short / put-side) and both booked
on a partial rather than a full target. The agent got there by a different route —
long first, released three minutes later on a premise fade, then reversed. That
reversal is what v4f's confirmation rule was written to produce, and the 10:07 exit
was a premise exit on a trade that was not yet losing, which is the disciplined
half of the rule v4h now bounds from the other side.

Two trades is a small sample and the basket was negative regardless. The narrow
signal worth keeping: across 12, 13 and 14 Aug every SL Hunting exit had a
nameable reason, and the two profitable days are the two where that held.

Also

• v4h wording assertions are now whitespace-insensitive. The prose is hard-wrapped,
  so a phrase can straddle a line break — the test was failing on a re-wrap that
  changed nothing about what the agent reads.
• Operational: the supervisor heartbeat from the session-state work is confirmed
  working in production — 65 heartbeat lines in the 14 Aug log. The 2026-08-12
  failure mode (snapshot writes stopping silently at 12:30 while trading continued
  to 15:10) would now surface within five minutes.

Gates

Gate	Result
Tests.test_nifty_multi_strategy_master	528 passed
Tests.test_market_data_health	28 passed
pytest "Tests/Signal Generators" "Tests/Dependencies" "Tests/Data Extractors"	1195 passed
Ruff	clean
mypy	54 files, clean
compileall	clean

🤖 Generated with Claude Code