cyb50-quant/research/dragon/v2/openspec/changes/followthrough-mid-exit-shadow-probe/design.md

Context

The prior shadow scaffold proved that delayed mid-zone reentry can happen, but the observed sample was sold at:

• 2020-12-09
• reason knife_take_profit_2_glued
• return about -0.24%

The same review also showed strong unrealized continuation after that exit. This suggests the engineering bottleneck may now be exit treatment, not signal discovery.

Goals / Non-Goals

Goals

• Keep RC1 default behavior unchanged.
• Add one narrow exit probe only for:
• entry code entry_glued_followthrough_reentry
• kdj_sell=True
• ql_sell=False
• mild positive repair state
• Make the result observable as a shadow branch.

Non-Goals

• Promote the branch to production alpha.
• Relax ql_rebound_weak_followthrough or high_zone_weak_b1.
• Rewrite the general glued sell family.
• Add multi-factor scoring or large experiment grids.

Decisions

1. Exit treatment is branch-gated and entry-specific

The new hold only applies when the entry itself is a glued followthrough reentry. Regular glued trades are not changed.

2. The hold only protects the first mild kdj_only wobble

The probe requires:

• kdj_sell=True
• ql_sell=False
• holding days still early
• a1 positive
• b1 only mildly weak
• c1 still in the mid-zone repair band

3. Observation is more important than immediate promotion

Even if the historical result looks strong, this stays a shadow branch because the sample count is still one repaired-trend path.

Risks / Trade-offs

• [Single-path evidence] accepted; this is exactly why the branch remains shadow-only.
• [Overfitting] mitigated by narrow scope and explicit shadow governance.
• [Default drift] mitigated by default-off config plus existing RC1 hash tests.

Migration Plan

1. Add config gates and exit helper.
2. Add branch config and daily-pipeline registration.
3. Add review script and fixture test.
4. Run review + full validation and publish acceptance summary.