[IMPROVEMENT] Mandatory small-slice decomposition before implementation selection

[acipat]

Feature Description

Problem Statement

Large issues consistently fail to merge because they are discovered mid-implementation to be bigger than estimated. The analysis stage already identifies decomposition candidates ("needs decomposition", "effort underestimated"), but this identification is advisory — implementation can still select an oversized issue and produce a half-finished PR that gets stuck.

Proposed Solution

Make the decomposition rule mandatory and blocking rather than advisory:

1. Any issue with estimated effort above a threshold (e.g., 0.4) MUST be split into independently mergeable child issues before the implementation stage can select it.
2. The analysis skill already flags decomposition candidates; this change makes the implementation skill refuse to select any issue flagged needs-split unless child issues exist.
3. The implementation skill checks for child issues before selection; if none exist, it skips the issue and logs "blocked: needs decomposition."

This is a policy change in the analysis/implementation skills, not new core code.

Value Proposition

• Impact: 0.8
• Effort: 0.3
• Priority Score: 1.41

Research Evidence

• Direct pipeline observation — multiple open issues flagged "needs decomposition" or "effort underestimated" in recent analysis cycles
• Half-finished PRs stuck in the merge queue

Implementation Plan

1. Add a blocking gate in the implementation skill: if the selected issue has label needs-split and no child issues exist, skip it.
2. Update the analysis skill to enforce that needs-split issues must have child issues created before they can be marked ready for implementation.
3. Add a validation check: the implementation skill queries gh issue list for child issues of the parent before selection.

Success Criteria

• No issue with effort >= 0.4 is selected for implementation without decomposition.
• Fewer half-finished PRs in the merge queue.
• Faster merge velocity for decomposed child issues.

[Da-Mikey]

Accepted — PR #598 implements mandatory decomposition gate. evolution_decomposition_gate.py blocks issues with effort_score >= 0.4 unless child issues exist, preventing monolithic picks from reaching PR stage.

[Lexus2016]

Blocked at integration review: PR #598 ships dead code.

What is wrong: scripts/evolution_decomposition_gate.py is implemented and tested, but nothing in the evolution pipeline actually calls it. skills/evolution/evolution-implementation/SKILL.md still handles size/needs-split manually and never invokes the gate, so the "mandatory" decomposition check cannot take effect.

Definition of done:

1. Add a concrete invocation in skills/evolution/evolution-implementation/SKILL.md at the pre-selection stage (before branching/implementation): run "python ${HERMES_HOME}/scripts/evolution_decomposition_gate.py check <issue_number> --effort \u003cfrom_analysis_json\u003e" and skip the issue when the script exits 1 (blocked: needs decomposition).
2. Make the gate read the effort_score from the analysis JSON so the skill does not have to hand-craft the --effort value.
3. Re-open a fresh PR referencing [IMPROVEMENT] Mandatory small-slice decomposition before implementation selection #579 once the gate is wired into the implementation skill and the full path is documented.