v0.13.2 — a refused manifest no longer costs the trace

Added

• smart_route() now names WHY it returned an empty menu: degraded_reason is http_<status> / timeout / circuit_open / transport, so a caller can tell a Cloud Run edge abort (a 0 s 429) from a slow read or an open breaker instead of filing every empty menu as "no skill offered".

• decimalai init --framework adk writes a runnable Google ADK agent:
  Gemini by default (gemini-3.5-flash; a non-Gemini --model is refused up
  front, because decimalai[adk] installs no other provider), a GOOGLE_API_KEY
  guard on line one, and skills delivered by prompt injection through
  instrument(agent_name=..., enable_skill_loader=True). The template was
  withheld on the premise that the hermetic J1 journey could not stand in for
  Gemini; that premise was false — google-genai honours GOOGLE_GEMINI_BASE_URL
  (since v1.72, google-adk 2.0.0's own floor), so the journey stub now speaks
  :generateContent and grades the ADK cell like the other three. Nothing in
  the generated file knows about the stub.

• Every generated file refuses to run without the provider key its MODEL
  needs, naming the variable and the Set: block decimalai init printed.
  Before, a missing OPENAI_API_KEY was a traceback from inside the provider
  client — after init(), instrument() and load_agent() had already made
  their network calls (on openai-agents, only at the first Runner.run_sync).
  The guard is derived from the same tables as the Set: block and is omitted
  where those tables only guess (a local provider, Pydantic AI's test model, an
  unknown prefix such as azure_openai:), so it never refuses a correct setup.

• The langchain template says at the call site that create_agent runs with
  recursion_limit=9_999 (not langchain-core's 25) and how to cap it once a tool
  is added — as visible as MAX_TURNS / MAX_REQUESTS on the other templates.

• The conformance probe serves a content_hash with every skill body, the way
  the platform does, and C13b now requires a model-pulled activation to carry
  that hash. Before, no conformance run ever saw a skill_hash, and the one
  fidelity clause that existed skipped a null silently — an adapter that stopped
  stamping versions would have stayed green.

Fixed

• A refused manifest registration no longer costs the trace. The caller-thread registration
  ladder is milliseconds by necessity, which cannot outlast a Cloud Run revision with no
  available instance; the trace was then shipped carrying a LOCAL id the platform had never
  stored and rejected with 400 manifest_id '...' does not exist. Production took 1,315 of
  exactly those in the 48 h to 2026-09-04, all from this SDK. Re-registration now happens on
  the background sender, where waiting costs the caller nothing: five jittered attempts,
  bounded, and the trace is held until an id exists rather than sent under one that does not.
  Applies to every adapter that registers a manifest — langchain, the generic
  @decimalai.trace path, ADK, LlamaIndex, OpenAI Agents and the Claude Agent SDK. The
  last four had the same defect with no retry ladder at all, in two shapes: two stamped
  the local id and took 400 manifest_id ... does not exist, two sent no id and took
  400 manifest_id is required. Both lose the trace.

• _poll_for_trace in the framework end-to-end tests raises TraceNeverArrived instead of a
  bare AssertionError whose message contained "Timed out" — the substring the live-test
  helper matches to classify a failure as provider-unavailable and SKIP it. A lost trace was
  being reported as a quota skip, the same defect class as the 2026-09-03 journeys finding.

• decimalai init gave up on the first admission abort. A single-instance
  backend at its concurrency limit answers HTTP 429 in ~0 ms with no Retry-After
  (Cloud Run's "no available instance"); measured on production 2026-09-03 that
  was 84-92% of requests in a ten-minute window. init made one
  /api/v1/auth/verify call — and decimalai init <agent> one bare
  GET /api/v1/agents and one GET /api/v1/agents/{name}/skills — and exited on
  any non-2xx, so the quickstart failed outright, and the scaffold wrote no
  agent.py, for anyone who ran it while the platform was busy (the fleet's
  scaffold canary filed those as decimalai_init_wrote_no_agent_file). All three
  calls now retry a 429/502/503/504 twice (0.5 s, then 1.5 s; a short
  Retry-After wins) before reporting the status. A 401/403 is a real answer and
  is never retried. The prompt read was already on the SDK's retry ladder and
  still only costs a note in the file, never the file.

• The openai-agents scaffold died with a traceback when the model called a
  tool the file does not define. The Support pack's starter prompt named
  kb_search; a model that called it raised
  agents.exceptions.ModelBehaviorError: Tool kb_search not found out of
  Runner.run_sync (the fleet's scaffold canary, 2026-09-04). run() now catches
  ModelBehaviorError the way it already caught MaxTurnsExceeded and returns a
  message naming the mismatch between the prompt's Tools line and the tools
  attached to the Agent. The hermetic J1 journey now replays that class: a second
  run of the generated file against a stub that calls an unknown tool must exit 0
  without a traceback.