gitreins-judge-criterion-scope-mismatch-false-fail

Answer

The solution is written to: ~/gitreins-judge-criterion-scope-mismatch-false-fail.md

What I did

I diagnosed the problem against the real tooling (not just the narrative) and verified every command:

• Installed gitreins 0.16.0 and read its actual judge/evaluator/pipeline/task-manager source (engine/evaluator.py, engine/pipeline.py, engine/task_manager.py).
• Reproduced the false FAIL and the fix end-to-end in a temp Go repo using gitreins' deterministic mock (GITREINS_MOCK_LLM_RESPONSE).
• Built and tested the pre-dispatch guard script embedded in the doc.

Root cause (short)

The criterion is the entire spec the tier-2 judge sees. A criterion transcribed from a board row's FIX text can demand production wiring, but the package is deliberately admission-only. Because the judge defaults to file_scope: changed and fast-track "do not trace into unchanged code", it cannot see the (unchanged) engine dispatch layer; the dead-code tools then report the symbol as "reachable only from the test". engine/pipeline.py:1053 turns any FAIL item into Overall: FAIL, and the foreman loop re-derives the same criterion.

Exact fix (verified)

# Diagnose
rg -n -i 'scope:|admission[- ]only|not wired|unwired' internal/<pkg>/
rg -n -g '!**/*_test.go' '\b<Symbol>\b' .     # only the definition => orphaned
git log --oneline --all --grep='DAGGER-049'    # siblings unwired too

# Fix: do NOT rework code — rewrite the criterion surface
gitreins task delete DAGGER-0504
gitreins task create DAGGER-0504 "<title>" \
  "At the <pkg> ADMISSION layer (this package is admission-only by design, not wired into the engine dispatch - the same scope as landed sibling rows <ids>), <observable admission behavior>"
gitreins task complete DAGGER-0504

# Optional prevention
boardctl update DAGGER-0504 --note "SCOPE: admission-only by design; criteria must name the ADMISSION layer"
scripts/gitreins-scope-check.sh internal/<pkg> <Symbol> -- <sibling-ids>

Verification results

The document contains the full root-cause analysis, the exact fix, the complete verified guard script, the verification transcript, the "when not to apply this fix" rule (to avoid rubber-stamping), and an appendix mapping each behavior to its source file/line.