fix(bus): check-stale-blockers re-flags a resolved stat-source PR citation forever

[asachs01]

Summary

check-stale-blockers flagged the same false positive on the same task twice, two days apart
(task_1786902033624, 2026-08-20 then 2026-08-22) — a real, correctly-cited PR (conduit#1424)
cited as the source of a supporting statistic ("documented in-code by conduit PR #1424 as 13
verified against prod"), not as the fix for the task's blocker. Both times a peer had to manually
re-trace and correct it, because the tool has no memory of a prior resolution.

This is a different shape from the existing precedent-citation false-positive class (2026-08-12,
PRECEDENT_CITATION_CUE_REGEX) — that class isn't naming a blocker at all ("same shape as the
action1 precedent, PR grandamenium#306"). This one names a real, relevant PR, just not as the fix.

Fix

Widens the existing exclude-list mechanism with a second cue category (documented ... by,
verified against, sourced from, per PR #NN's own data/finding/...) rather than adding new
persisted state (no resolved-citations file to keep in sync). Reuses the exact regex-window
technique already proven in production for the precedent-citation class.

Known limitation, found and documented rather than silently fixed

Writing the "mixed citation" regression test surfaced a pre-existing architecture gap: a cue
phrase sitting in the textual GAP between two close PR references can leak into the next
reference's before-window, because the existing prev-match-end clamp only protects against
reaching before/into the previous match's own span, not the prose after it. Not fixed here (out
of scope for this diff) — documented in the test file as a known limitation for follow-up if a
real case is ever reported. Scoped test uses well-separated sentences to avoid it.

Verification

• All done in an isolated git worktree.
• New/existing targeted tests: 48/48 pass (tests/unit/bus/system.test.ts), including 3 new
  cases for the real incident text, a mixed-citation case, and two more cue phrasings.
• Broader run: tests/unit/bus/ + tests/integration/: 857 passed, 3 failed — all 3 are
  phase5-performance.test.ts cron-startup timing thresholds, already documented in this repo's
  own CLAUDE.md as environment-flaky under load, unrelated to this change (confirmed: no code
  touched here is anywhere near cron-startup timing).
• npm run build succeeds.

Not merging/self-approving

Opened for review per normal process.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is enabled.}