bug: saved FixMap reports rank as high-confidence changed files in later --working-tree maps, contradicting the output-file exclusion guarantee

[aryamthecodebreaker]

Summary

The docs promise:

The current issue, comparison, verification, and output files are kept out of ranking, change detection, and cache state, so a saved FixMap report cannot recommend itself.

That exclusion appears to apply only to files written by the current invocation. A report saved by a previous invocation ranks as a high-confidence "changed file" in later plan --working-tree runs, appears under Changed Files, and outranks task-relevant ranking candidates. fixmap setup artifacts behave the same way.

Environment

• fixmap 0.9.0 (npm global)
• Windows (win32), Node v24.13.0

Repro

git init demo && cd demo
Set-Content package.json '{"name":"demo","scripts":{"test":"node --test"}}'
New-Item src, test | Out-Null
Set-Content src/pay.js 'export function pay(amount){ return amount*2; }'
Set-Content test/pay.test.js 'import "../src/pay.js";'
git add -A; git commit -m init

# 1. Save a report at the repo root (a very common workflow)
fixmap plan --issue "fix pay" --format json --output saved-report.json

# 2. Edit the source (uncommitted)
Set-Content src/pay.js 'export function pay(amount){ return amount*3; }'

# 3. Map the working tree
fixmap plan --working-tree --include-untracked

Actual (step 3)

## Context Files

- `src/pay.js` (high confidence, score 57): changed file; ...
- `saved-report.json` (high confidence, score 22): changed file; content matches task terms: pay

...

## Changed Files

- `saved-report.json`
- `src/pay.js`

A stale FixMap artifact is presented to the agent as a high-confidence inspection target and as a changed file of this edit. In sessions with several accumulated reports (out-1.json, out-2.json, …) they crowd the default --limit 8: in one observed run, six old reports plus a scratch JSON filled seven of eight context slots while the only relevant source file ranked behind them.

Related downstream effect seen in that crowded run: the Impact Graph routed the repo's test onto an artifact instead of the source file it imports (test/auth.test.js … routed test for big-plan.json). That attribution bug has a separate, cleaner repro (issue to follow), but artifact crowding is what exposed it.

fixmap setup outputs (.claude/skills/fixmap/SKILL.md, .cursor/commands/fixmap.md, …) rank identically as high-confidence changed files after any working-tree edit, since most users don't commit them immediately.

Expected

FixMap-generated artifacts should not compete with real source in later maps — per the quoted guarantee, output/report files should be kept out of ranking and change detection regardless of which invocation produced them, or at minimum be heavily deprioritized in --working-tree mode.

Suggested fix

FixMap already ships a structural validator (fixmap validate). When scanning untracked/changed paths, parse each candidate JSON at the repo root: if it validates as a FixMap report (or carries the reportVersion marker), exclude it from ranking/change detection the same way the current invocation's --output target is excluded. Setup artifacts could be recognized by their fixed paths (.claude/skills/fixmap/**, .cursor/commands/fixmap.md, .github/prompts/fixmap.prompt.md, .agents/skills/fixmap/**).