Proof output retention Phase 4: oracle evidence alignment

[tonyketcham]

Summary

Give oracle tasks the same explicit split between display tails and full execution evidence.

Source Context

• Cursor cloud-agent follow-up: https://cursor.com/agents/bc-51322132-f3a3-4f80-b0cd-0ed319ccce62
• Implementation PR: docs: plan proof output retention #199
• Proposal artifact: https://github.com/FlatbreadLabs/flatbread/blob/toeknee/proof-output-retention-ce62/docs/proposals/proof-output-retention-plan.md
• Review artifact: https://github.com/FlatbreadLabs/flatbread/blob/toeknee/proof-output-retention-ce62/docs/proposals/proof-output-retention-review.md
• Judge artifact: https://github.com/FlatbreadLabs/flatbread/blob/toeknee/proof-output-retention-ce62/docs/proposals/proof-output-retention-judge.md
• Phase-specific artifact section: https://github.com/FlatbreadLabs/flatbread/blob/toeknee/proof-output-retention-ce62/docs/proposals/proof-output-retention-plan.md#L281-L294

Why This Exists

• Phase 1 intentionally scopes execution-authoritative transcript behavior to kind: 'task'; the README/SKILL carve out oracle evidence as still tail-bounded.
• Today oracle stdout/stderr are formatted into resultText using the display tail cap, with no full-evidence artifact equivalent to task stream transcripts.
• Phase 4 aligns oracle output with the broader execution-vs-display model while preserving the existing ## Stdout (tail) and ## Stderr (tail) display shape.

Acceptance Criteria

• Write ${taskId}.stdout.log and ${taskId}.stderr.log under the artifact directory for oracle tasks when artifacts are enabled.
• Add a ## Evidence footer and canvas pointers for stdout/stderr paths without bloating STATE.tasks[].resultText.
• Apply --max-in-memory-output-bytes to oracle full evidence in --no-artifacts mode, producing BUDGET-EXCEEDED on overflow.
• Ensure downstream tasks depending on oracles use outputPolicy.upstream for long oracle evidence rather than silently seeing only the legacy tail.
• Cover full stdout persistence, no-artifacts overflow, and downstream policy-bounded oracle context with tests.

Verification

• pnpm -F @flatbread/proof test
• pnpm -F @flatbread/proof typecheck
• pnpm verify

Related Phase Issues

• Proof output retention: Phase 0/1 follow-up hardening #200 — Proof output retention: Phase 0/1 follow-up hardening: Proof output retention: Phase 0/1 follow-up hardening #200
• Proof output retention Phase 2: upstream prompt policy and budget preflight #201 — Proof output retention Phase 2: upstream prompt policy and budget preflight: Proof output retention Phase 2: upstream prompt policy and budget preflight #201
• Proof output retention Phase 3: resume, supervisor, and disk ergonomics #202 — Proof output retention Phase 3: resume, supervisor, and disk ergonomics: Proof output retention Phase 3: resume, supervisor, and disk ergonomics #202
• Proof output retention Phase 4: oracle evidence alignment #203 — Proof output retention Phase 4: oracle evidence alignment: Proof output retention Phase 4: oracle evidence alignment #203
• Proof output retention Phase 5: documentation and skill refresh #204 — Proof output retention Phase 5: documentation and skill refresh: Proof output retention Phase 5: documentation and skill refresh #204

Artifact Linkage

This issue is linked back from docs/proposals/proof-output-retention-plan.md and, where relevant, docs/proposals/proof-output-retention-review.md so the repo artifacts and GitHub issues remain navigable in both directions.