cyb50-quant/research/dragon/v2/openspec/changes/glued-followthrough-shadow-reentry/design.md

Context

RC1 currently uses buy_block_glued_high_weak_rebound as a quality-preserving hard veto. Historical evidence does not support globally relaxing this block, especially for ql_rebound_weak_followthrough.

The requirement is to add a minimal, isolated, reversible research path:

• day-0 remains blocked
• selected subtype can be queued as pending in a probe branch
• delayed reentry is allowed only after explicit confirmation

Goals / Non-Goals

Goals

• Preserve RC1 default branch behavior exactly.
• Keep buy_block_glued_high_weak_rebound hard block in default path.
• Keep ql_rebound_weak_followthrough hard-blocked by default.
• Add a branch-gated pending scaffold for followthrough research.
• Make behavior locally testable.

Non-Goals

• Promote probe behavior to formal alpha.
• Relax global veto in RC1.
• Rework sell families in this phase.
• Introduce a broad scoring framework.

Decisions

1. Base path hard-veto remains unchanged

Blocked signals still return buy_block_glued_high_weak_rebound on the current bar. Pending only records a candidate in probe branch.

2. Phase-1 pending subtype scope is narrow

Only mid_zone_very_weak_b1 is enabled in the probe branch by default.
high_zone_weak_b1 and ql_rebound_weak_followthrough remain disabled.

3. Confirmation requires explicit repair

Probe reentry requires all of:

• at least one bar waited
• no sell cross
• ql_buy=True
• close > signal_close
• b1 - signal_b1 >= threshold

4. Keep pending implementation isolated

New glued pending logic is isolated in a dedicated module and dedicated strategy-context fields; no merge with deep-oversold module.

Risks / Trade-offs

• [No immediate performance gain] acceptable; this phase targets structure and observability first.
• [Default-path drift] mitigated by hash regression and golden regression tests.
• [Probe path overreach] mitigated by strict branch-level config gating.

Migration Plan

1. Add config flags + probe branch config.
2. Add glued followthrough confirmation module.
3. Wire pending lifecycle in strategy facade (queue/confirm/clear).
4. Register new reason family mapping.
5. Add shadow branch to daily pipeline list.
6. Add tests and run full validation.

Rollback Plan

If any RC1 default regression appears, revert:

• probe branch config entry
• glued pending fields and strategy hooks
• glued followthrough module
• daily pipeline shadow branch registration

RC1 default behavior must remain unchanged after rollback.