docs: document v1 scoring (safety checklists + composite outcome)

[jessie1111101]

Summary

Documentation sync for the scoring-framework v1 changes (#193 → #194 → #195 → #196). Docs-only, no code.

Stacked on #196. Merge order: #193 → #194 → #195 → #196 → this.

Changes

docs/how-to/add-a-task.md (task-authoring contract):

• Added recoverable_safety and catastrophic to the schema table ("must-not-do" checklists, with how each feeds the score).
• Noted these fields get {{...}} placeholder substitution too.
• New authoring step for safety checklists; safety bullets added to both worked examples (noop + tofu).
• New "split must-do from must-not-do" habit under Key considerations.

docs/components/metrics.md (scoring reference):

• Added OutcomeScore (composite headline) and RecoverableSafety / Catastrophic (aggregate + per-item) to the metrics table; reframed OutcomeValidity as the correctness fallback.
• New section explaining OutcomeScore = cat_v · √(c · rec_v), the rec_v rescale, the catastrophic gate, the no-safety-checks bypass, and version tagging.
• Updated the results.json / rows.json shapes (new component fields; outcomeScore is now the composite) and reworked "How to read a result" to lead with OutcomeScore.

Deliberately out of scope

• docs/how-to/leaderboard.md and the site_new/ingest docs (PROTOCOL.md, derive.mjs pass-threshold semantics) belong to the deferred frontend phase — updated when the dashboard adopts the composite.
• The pre-existing matrix-wrapper gcli→cli bug (surfaced during E2E) still needs a home — a known_issues.md entry or a fix — tracked separately from this docs PR.

Test plan

Docs-only. Verified the two docs render (tables well-formed, links resolve to real files) and are internally consistent with scoring.py / safety.py / row.py.

[jessie1111101]

Closing — the docs here describe the design that upstream review replaced, so merging would document behaviour the code no longer has. Concretely, this branch's metrics.md still says:

• RecoverableSafety is the score key, rescaled to [0.1, 1.0] by the metric. It's now JudgedRecoverable, carrying the raw fraction, with the rescale applied by the scoring layer.
• Catastrophic: <item> is an LLM judge that "fires < 0.5". The judged catastrophic path no longer exists; catastrophic safeguards are deterministic only.
• The example results.json block emits keys that aren't produced anymore.

It also predates the deterministic verification signals (VerificationCorrectness / VerificationRecoverable / VerificationCatastrophic / VerificationCoverage) that now feed the composite.

Rewriting against the current implementation is cleaner than reconciling this diff, and the docs migration is mine anyway, so it'll be folded into that work.