TASK: Exit-geometry interim fix — measure the real MFE distribution from
      existing position telemetry, then bring runner_target_atr_mult down
      from fantasy (6.07 R) to where this strategy's moves actually live.
      ONE variable. Runs AFTER task_phase0a_excursion_columns is green
      (it consumes the backfilled columns).

## Established context

- Root-cause assessment 2026-09-06 Finding 2: the plan promises R:R 6.07:1
  (runner_target_atr_mult 7.59 ÷ option_stop_atr_mult 1.25) while
  trailing_stop.v3 activates at +1.0 R and trails 1.0 R behind peak.
  Reaching the target requires a 6.07 R advance with no 1 R retrace.
  Result across 31 Aug–4 Sep: 0 targets in 21 trades; target_exit.v1 is
  dead code; realized payoff capped ~1.3:1 against a 44% break-even win
  rate the entry cannot deliver.
- Interim MFE evidence from the EDIT/peak logs (9 trailed trades): peaks
  in R = 1.04, 1.04, 1.16, 1.19, 1.26, 1.46, 1.95, 2.09, 2.10. The other
  12 trades peaked below 1.0 R. Maximum observed excursion ≈ 2.1 R.
- 7.59 was declared by the 2026-08-14 amendment as a runner ceiling, not
  a tuned value. Replacing an untuned constant with a MEASURED one, with
  the measurement printed in the amendment, is the same class of act as
  the premium floor — geometry arithmetic, not outcome-chasing (RR-2).

## IMPORTANT FINDING — measure first, then set, in the same task

Phase 0a has just backfilled mfe_pct/mae_pct for every closed position
from OBSERVED telemetry. This task's first step is the measurement:

  The MFE-in-R distribution of ALL closed atr_derived-era positions
  (1 Sep onward; exclude the premium_pct-era and 0-DTE trades, stating
  the exclusion): per-trade MFE ÷ R table, then p50/p75/p90/max.

The new target multiple = the value that puts the target at
approximately the p75–p90 of WINNING excursions — reachable by the moves
this strategy actually produces, so target_exit.v1 comes back to life —
expressed in ATR terms (target_R × 1.25). From the interim numbers that
lands near 2.0–2.5 R (runner_target_atr_mult ≈ 2.5–3.1); the agent's
measured table decides the exact value, and the amendment prints the
table it was read from. If the measured distribution contradicts the
interim table above (e.g. p90 of winners ≥ 3 R), follow the measurement,
not this prompt.

## Verify the premises

1. Phase 0a's columns are filled for the atr_derived-era trades (this
   task's input; if 0a has not run, STOP).
2. Every consumer of runner_target_atr_mult / the plan target: entry_plan,
   evidence rows, admission, any R:R validation — especially whether
   strategy_registry's validated combination (1:2) or any gate ASSERTS a
   planned R:R value. If something pins 6.07:1 or validates the target
   multiple, list it — the change must keep every gate green honestly,
   not by weakening an assertion.
3. The three lockstep config artifacts for the key (spec / schema.py /
   seed SQL) and its schema range — the new value must sit inside the
   declared range, or the range moves in the same amendment.
4. trailing_stop.v3, break_even.v2, activation/trail multiples: UNTOUCHED.
   Confirm no test derives trail behaviour from the target value.

## Do NOT change

option_stop_atr_mult (1.25) · trailing_stop.v3's activation and distance ·
break_even, partial_exit, spike_exit, session_close · the stop model ·
any filter or threshold · ai.confidence_threshold — ONE variable changes:
the target multiple. The wider exit-geometry question (activation level,
trail distance, small-profit capture below 1 R) is EXPLICITLY DEFERRED to
the campaign after ~2 weeks of F-7 data — record it as deferred in the
amendment so the decision is dated.

## Implementation

1. The measurement (table + percentiles) — computed and printed FIRST.
2. Spec 05 amendment (dated): the measured table, the chosen multiple,
   the reasoning (target at p75–p90 of winning MFE; former 7.59 ceiling
   unreachable by construction against the 1.0 R trail — cite the
   assessment doc), and the INTERIM label: to be revisited by the
   exit-geometry campaign once the F-7 corpus reaches ~50 trades.
3. The three config artifacts in lockstep.
4. Nothing else.

## Tests

- Config sync gate green; boot invariantry unchanged.
- entry_plan produces the new target on a fixture; geometry gate verdict
  unchanged (it reads the stop, not the target — verify, don't assume).
- Any R:R-printing surface shows the new planned ratio consistently.
Then the FULL suite:

    bash backend/scripts/ci/run_all.sh

All 13 gates green.

## Verification report

A. The measured MFE table + percentiles  B. The chosen multiple and its
arithmetic  C. Premise-2 consumer inventory (anything that pinned 6.07)
D. Spec/schema/seed diff  E. Old vs new cfg- checksum  F. 13 gates
G. Premises that did not match  H. Unknowns

CRITICAL: ONE variable. If premise 2 finds a consumer that VALIDATES the
old ratio (not merely displays it), STOP and report before changing it —
that consumer is a design statement someone made on purpose, and Nitin
rules on it.