feat(router): preserve Auto policy authority and response evidence

[drewstone]

Change

Support inline Router Auto policies through the existing router-tools executor.
Check strategy and assessment models against the profile and allowed model set.
Keep every buffered HTTP response, including errors, in native transcript capture.
Provide an awaited onProviderResponse observer for durable live capture.
Retain the raw response and leave aggregate Auto usage unknown.

Actual consumer proof

All coding and verification ran in Tangle Sandboxes; GTR only operated the SDK.

• Published Runtime0.289.1 failed on the real Django development task: served claude-sonnet-5 versus requested tangle/auto.
• Patched packed consumer completed the same real task: Sandbox sandbox-837da58bd4a2, PID207, three Auto calls, read_file and edit_file, 1113-byte source patch.
• Independent frozen repo checker in a separate Sandbox accepted the model patch: exit0, resolved, full class and regression checks exit0. This is not an official SWE-bench score.
• Final-source Auto pilot in sandbox-b043498d207b exercised the native Jev selector. The cheap generation was refused for a Chat Completions reasoning/tools limitation. The complete400 response retains the paid selector receipt and failed generation ID. This failed pilot is not a success claim.
• One unique development source unit across these attempts. No confirmation split opened; no economic or >=200-unit campaign claim.

Private complete records: /mnt/traces/autoroute-executor-20261001 on drew-gtr-pro.
Lab issue1115 remains open.

Engineering checks

Typecheck, example typecheck, Biome, build, maintained model execution boundary, API surface and docs freshness passed.
Generated docs and minor release note included.
No unit tests added or run.
No version allocation; maintained release conductor owns publishing.

Risk / rollback

Additive observer and evidence fields.
Auto requires a resolved inline policy and buffered transport.
Observer failures stop execution; concrete model identity checks remain strict.
Revert this change to restore previous behavior.

[tangletools]

✅ Auto-approved PR — 29e56df0

Blanket team auto-approval is intentional. The merge gates are CI and codex-p1.
No automated review runs on this PR. This approval rests on the rule above alone.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-10-01T12:29:35Z}

[tangletools]

✅ Auto-approved PR — 078de78d

Blanket team auto-approval is intentional. The merge gates are CI and codex-p1.
No automated review runs on this PR. This approval rests on the rule above alone.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-10-01T12:44:34Z}

[drewstone]

Independent current-head review passed078de78df5fe1152035f81a79afb10ae966d8db4. The two review findings are fixed: evidence-observer failures remain terminal before timeout/network retry classification, and Auto dollar subtotals remain unknown aggregate usage. Actual paid HTTP checks for retryable SDKError, fetch TypeError and deadline stall each retain one200 response and stop with one request. I independently matched the latest candidate patch SHA256 to its separate frozen checker receipt (exit0/resolved) and counted three logical responses/eleven physical call receipts. Declared concrete model authority remains checked; exact Router policy digest validation belongs to the frozen consumer and is documented. This is one development source unit, not confirmation or causal superiority. Receipt /mnt/traces/discovery-issue-review-20261001/runtime1489-independent-review.json.