fix(outcome-collector): update report for new Safe Output Outcome Evaluation

[Copilot]

The outcome-collector workflow had stale paths and a report format that predated the typed outcome evaluators introduced in ADR-35218 (add_reviewer, submit_pull_request_review, update_issue/update_pull_request with evidence strength, ignored state, fallback_exists_only signal).

Path fixes

• Telemetry export never ran: Export outcome telemetry checked /tmp/gh-aw/agent/outcome-evaluations.jsonl but evaluate_outcomes.cjs writes to /tmp/gh-aw/outcome-evaluations.jsonl — telemetry was silently skipped every cycle
• Agent read paths: Removed spurious /agent/ prefix from outcome-summary.json and outcomes/run-*.json references in the markdown body

Report format updates

Scorecard — added Ignored row; expanded Accepted into evidence-strength sub-rows:

| Accepted        | 42 / 60  | —                          |
| — strong        | 28       | merged, completed, approved |
| — medium        | 11       | engaged, retained           |
| — weak          | 3        | existence only              |
| Rejected        | 8        |                            |
| Ignored         | 10       | no observable follow-up     |

Median resolution — field is median_resolution_sec (int, seconds); scorecard now shows the conversion formula inline (÷ 3600 → hours, ÷ 60 → minutes)

Evidence Quality section — surfaces fallback_exists_only_count with context: items evaluated via the generic existence fallback inflate accepted_weak and may overstate acceptance

Action items — two new checks: ignored rate >30% (targeting/quality signal) and fallback evaluation rate >20% (data quality signal)

Per-workflow table — added Ignored column; removed Reactions (not present in summary JSON)

Field reference table — documents all summary JSON fields produced by evaluate_outcomes.cjs, including accepted_strong/medium/weak, fallback_exists_only_count, ignored, and noop