Portable Claude profile: no silent fallback to an older model; StructuredOutput fields at the top level - #272

Scope

• What this PR changes: the portable Claude profile stops the silent fallback to an older model and asks workers to return StructuredOutput fields at the top level of the call arguments. The dated record is docs/decisions/2026-09-25-model-fallback-guard.md.
• Base commit: 5abd17e4 (origin/main, which includes Token-efficiency coverage check: adoption profile plus value-free --client-wiring for every host #264 and Hugging Face sign-in as a value-free native credential: status rows, guard, deny rules, per-host token policy #268)
• Lane: lane:foundation
• Owned paths touched:
  • adoption/templates/claude.settings.template.json: env CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1 and "switchModelsOnFlag": false, in hunks separate from Hugging Face sign-in as a value-free native credential: status rows, guard, deny rules, per-host token policy #268's deny-rule hunks
  • examples/claude-native/CLAUDE.md: two sentences. One puts StructuredOutput fields at the top level. The other says what to do when a worker returns null: check the transcript for a safety refusal, then edit the brief and retry on the current model, and never accept an older model's answer.
  • docs/decisions/2026-09-25-model-fallback-guard.md (new)
  • hot file: manifests/evidence.json, re-registering the two changed hashes in the last commit

Evidence-class table

Local commands run

$ python3 -m unittest tests.test_credential_status tests.test_install_claude_profile tests.test_render_config tests.test_apply_claude_settings tests.test_adjudicate tests.test_claude_lane tests.test_credential_tools tests.test_landscape tests.test_verdict_lane_vendoring tests.test_record_verdicts
Ran 486 tests ... OK
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 5601, "profiles": 4, "receipts": 147, "status": "passed"}
$ python3 scripts/component_matrix.py --check; python3 scripts/new_host_grand_list.py --check
checked; passed
$ gitleaks git --log-opts="origin/main..HEAD" --redact .   (guarded ecosystem wrapper)
2 commits scanned; no leaks found

Independent review

An evidence-reviewer agent (Opus 5.5 at max) reviewed six rounds. It checked each fix against the 2.1.282 binary, the live docs, the catalog and GitHub.

• Rounds 1–5: needs_changes. All findings are resolved. They included:
  • the opt-out on an already-guarded host;
  • bio → Opus 5;
  • the silent lane being server-armed;
  • which agents omit CLAUDE.md and how the rule can still reach them (agent body, SubagentStart hook, brief);
  • the population and cutoff of the source-scout figure;
  • the /config row;
  • the status of the Token-efficiency coverage check: adoption profile plus value-free --client-wiring for every host #264 watch.
• Round 6: agree. What stays unproven, by design: the refusal behaviour itself, which is source_review.

Decision record

docs/decisions/2026-09-25-model-fallback-guard.md. It names the evidence, the alternatives (keep the switch; the flag alone; an availableModels allowlist; agent bodies only) and the overturn conditions (a documented subagent-wide control, a client that stops reading the variable, which agent-lab's release-watch probe reports with exit 3, or a same-version A/B showing no StructuredOutput gain). The fallback variable is undocumented, so that part stays keep-but-compare.

Checklist

• No workflows changed.
• No secrets printed, logged or committed.
• No paid hosting or billing surface.
• Peer-owned untracked files and worktrees were preserved. The adoption/agents/claude StructuredOutput line stays on the rollout's contract track.

🤖 Generated with Claude Code