[Task]: wingetbot worker — Internal/infra-error class learnings

Nobody has claimed this yet.

### Description

Per-class learning tracker for the **wingetbot auto-update PR triage worker** (#411124). Capture evidence and heuristics for the **Internal / infra error** class here.

**Labels:** `Internal-Error`, `Internal-Error-Dynamic-Scan`, `Internal-Error-Static-Scan`, `Internal-Error-Domain`

### Current disposition (from the 2026-08-01 pilot)

**Recommend retry — with a ceiling.** Transient pipeline faults are fixed by a moderator `@wingetbot run`. The worker **recommends** the retry; it never posts the trigger itself.

### Learnings so far

- Example #404062 (OSGeo.QGIS): wingetbot had **already retried `@wingetbot run` 6×**, then a moderator (`stephengillie`) posted the deterministic-failure notice → retries were exhausted and a human took over.
- **Retry has a ceiling:** count prior `@wingetbot run` comments; once exhausted and/or a moderator has already engaged, **escalate/defer instead of recommending yet another retry**.
- The NIUBI.PartitionEditor set (#410733–#410737) shows infra errors can hit a batch of related PRs at once.

### Open questions / learnings to capture

- [ ] What retry ceiling is right (3? 5?) before switching from "retry" to "escalate"?
- [ ] Reliable parse of prior `@wingetbot run` count + whether a moderator has commented (reuse the general defer rule).
- [ ] Distinguish truly transient infra errors from deterministic failures mislabeled as internal.

Parent: #411124

_Created with GitHub Copilot assistance._

Start with parent #411124 and review examples #404062 and #410733–#410737. Determine a retry ceiling, define how prior @wingetbot run comments and moderator involvement are parsed, and document how deterministic failures are distinguished from transient infrastructure errors. Done means the retry-versus-escalate rules and evidence are settled.

Written by the indexing model from the issue text.