[rank] Explicitly naming a dist/ generated path still grants high confidence

[aryamthecodebreaker]

Summary

Mentioning packages/action/dist/index.mjs in task text marks it explicitly named and, as the leader, confidence: high — even though a maintained source counterpart exists under packages/action/src/. Generated-duplicate filtering is skipped for mentions.

Evidence (v0.8.1)

fixmap plan --issue https://github.com/aryamthecodebreaker/FixMap/pull/1 ranked packages/action/dist/index.mjs #1 high with reason explicitly named in the task.

Expected

When a mentioned path is generated and a maintained stem twin exists, prefer/redirect to the source (or cap confidence and say the mention is a build artifact).

[aryamthecodebreaker]

Attempted twice for v0.8.2 and reverted both times. Staying open.

The failure is real

$ fixmap plan --issue "the bundle in packages/action/dist/index.mjs mishandles the comment marker"
  1. packages/action/dist/index.mjs [high]  explicitly named in the task

A named path bypasses generated-duplicate filtering (that is #371), so quoting a stack frame or a build artifact puts build output at rank 1 labeled high.

Attempt 1 — redirect the mention to the maintained source

Worked on the symptom (packages/action/src/index.ts took rank 1) but broke three tests, including one that has asserted the opposite since the generated-duplicate rule was written:

rankContextFiles > still ranks a build output copy that the task names explicitly
task text: "dist/color-support.js is stale and disagrees with the source"

That is a task genuinely about the artifact, and redirecting to source loses the point entirely.

Attempt 2 — add the source beside the mention instead of replacing it

  1. packages/action/dist/index.mjs [high]
  2. packages/action/src/index.ts   [medium]

The source surfaces, but the artifact still leads at high — so the actual complaint in this issue is unaddressed — and 2 tests still failed.

Where it stands

Held-out and regression are unchanged under both attempts, so the suites cannot distinguish any of the three behaviors. The tension is genuine: "the built file is wrong" and "the behavior in this module is wrong, and I happened to quote the built file" are different tasks that look identical to the ranker.

The distinguishing signal is probably in the task text — words about staleness and artifacts versus words about behavior — but that is a new classifier with no evidence behind it, and shipping unmeasurable ranking changes is what #274 was about.

A narrower option not yet tried: leave the ordering alone and only cap confidence on a generated file that has a maintained twin, with a reason naming it as build output. That addresses "grants high confidence" without reordering anything. Worth doing when there is a case that can score it.