Two issues caught while live-validating v1.5-C against real
providers:

1. **Stale skill prompt.** The S1 driver's ``responder`` skill was
   written in the Phase 15 (response_format JSON) era; its
   system_prompt told the LLM "respond in one sentence" with no
   markdown contract instructions. Phase 22 (markdown-primary turn
   output) made that fail with ``EnvelopeMissingError`` because the
   parser has nothing to lift. Add the
   ``## Response`` / ``## Confidence`` / ``## Signal`` contract
   block to the prompt — same pattern as the production skill
   prompts under ``examples/incident_management/skills/*/system.md``.

2. **No Azure parametrize arm.** The driver covered ``workhorse``
   (OpenRouter) + ``local`` (Ollama). Azure has been first-class in
   ``runtime.llm.get_llm`` since Phase 13 but had no live verification
   path. Add an ``azure`` arm parametrize that constructs an
   ``AzureChatOpenAI`` from ``AZURE_OPENAI_KEY`` + ``AZURE_ENDPOINT``
   + ``AZURE_DEPLOYMENT`` (defaults to ``gpt-4o``).

Per-leg skip semantics: each arm independently skips when its keys
are absent. Replaces the global ``pytestmark.skipif`` that required
ALL three keys for any leg to run — partial-key environments now
exercise whichever providers they can reach. Drops the
``_OPENROUTER_KEY and _OLLAMA_KEY and _OLLAMA_BASE_URL`` global
gate; the per-leg gate inside the test body owns it.

The ``LLMConfig`` builder also handles a fully-keyless environment
by falling through to a stub provider so config validation passes
during test collection.

Live verification status (with the keys in this dev environment):
  * ``local`` — PASSES against Ollama Cloud gpt-oss:20b
  * ``workhorse`` — fails on credit / rate-limit (account-specific)
  * ``azure`` — fails on connection error (placeholder endpoint in
    .env; framework path itself is intact)

…6-1t:free

Demonstrates the v1.5-C per-agent provider story end-to-end with
two REAL providers in flight:
  * intake (skill override)        → Ollama Cloud gpt-oss:20b
  * triage / DI / resolution (default) → OpenRouter inclusionai/ring-2.6-1t:free

The free OpenRouter tier rate-limits aggressively; the
preceding ``feat(retry)`` commit's 429 backoff (7.5s/15s/22.5s)
keeps multi-agent INC runs working through transient throttles.

Operators on a paid OpenRouter plan should swap the model back to
``openai/gpt-4o-mini`` (or any other paid model) — the rest of
the registry is unchanged.

Bundles dist/app.py + dist/apps/{code-review,incident-management}.py
in line with the ``runtime.graph._RATE_LIMIT_MARKERS`` +
``_ainvoke_with_retry`` rate-limit branch from the preceding feat
commit. No bundle-only edits.