feat(papercuts): /defect-session papercuts dispatches fixes from clusters

[nohat]

Part of #22. Wave 4. Blocked on #30. #31 is a soft dependency (the repro kit).

Problem

Triage produces issues; fixing them is dispatched by me with /defect-session. The session has no way to say "work the papercut clusters", and nothing makes the research-backed fix policy explicit for these reports.

Plan

1. A papercuts mode for /defect-session (.agents/skills/defect-session/): lists clusters whose issues are open, ordered by count times recency times the issue's Severity: line (a sort, not a model), shows the top few, and runs the existing loop on the ones I pick. Evidence comes from t3 papercuts show and the repro kit.
2. Reproduce first: a failing test (unit level is fine) built from the event buffer and trace slice, or the repro kit for UI cases. No failing test means a diagnosis-only change or a bounded mitigation, with the issue staying "mitigated, cause unknown" as the charter already says.
3. Scope: prefer single-file, bounded fixes. A fix that spans files is flagged as large in the issue and stays human-led. Published results show success dropping sharply on multi-file and long patches; treat that as a prior, not a rule.
4. Review: the fresh reviewer also tries to disprove the failing test, not just the fix (a reproducer can capture only one manifestation).
5. Attempt budget: at most 2 attempts per cluster per day, then a needs-human line on the issue.
6. After deploy: t3 papercuts mark <id> fixed --fixed-in <fork version> for every report in the cluster and a comment on the issue with the version; this starts the soak (feat(papercuts): recurrence per build, soak, and automatic close #33).
7. Dispatch stays mine. Nothing auto-starts. The triage digest names the command to run. Whether to auto-dispatch for a class of cluster with a good track record is deferred until the metrics in feat(papercuts): recurrence per build, soak, and automatic close #33 exist.

Acceptance

• The skill's papercut mode picks a cluster, fetches evidence through the CLI, and leaves the issue in the right state for each outcome (fixed, mitigated, needs-human).
• A fix to a seeded known cluster goes through the gate and ends with the overlay and issue updated.

Surfaces

Skill and docs only. Update docs/fork/defect-resolution.md ("The loop here", Discover) to name the papercut source instead of "once they exist".

Related

#22, #30, #31, #33.