Wire tier-1 typed-event conformance fixtures

[chris-colinsky]

Summary

First tier of the conformance-harness fixture catch-up: wire the typed
LLM-event observability fixtures that python currently satisfies through its own
unit tests into the YAML conformance harness, so the harness runs the spec's own
fixtures rather than python-authored equivalents. Test-only, no library change,
no pin bump; this strengthens the cross-impl conformance signal.

Wired (10)

Moved from _UNIT_TESTED_FIXTURES to _SUPPORTED_FIXTURES:

• 060-065, 068: LlmCompletionEvent field population (input messages, output
  content, request params and extras, active prompt plus the null case,
  response model)
• 067, 071: call_id distinctness across a multi-node chain
• 072: LlmFailedEvent / LlmCompletionEvent mutual exclusion

Harness machinery added

• A multi-node-chain runner (_build_chain_llm_graph) for the chained-call
  fixtures, alongside the existing single-node typed-event runner.
• New assertion shapes: fields_absent_keys, the event_counts list form, the
  non-empty <any-string> matcher, and a recursive value matcher so
  active_prompt (a PromptResult) compares against the fixture's identity
  mapping.
• renders_prompt active-prompt binding, request-model binding from
  calls_llm.model, and RuntimeConfig construction from calls_llm.config.
• expected_error now asserts both the error category and the originating node;
  the call_id presence/distinctness invariants for 067/071 are machine-checked.
• Consolidated the mock-transport and RuntimeConfig construction into shared
  helpers.

Deferred (4)

Four fixtures stay in _UNIT_TESTED_FIXTURES with documented reasons, each
blocked on an upstream spec fixture change to be picked up at the next pin bump.
The behavior is already covered by the unit suite, and the coverage guard keeps
the accounting honest:

• 066: the prompt group has a single member at the current pin; python's
  PromptGroup requires at least two. The corrected fixture is upstream.
• 069: asserts a request model the fixture does not declare.
• 070: the malformed tool message (no tool_call_id) is non-constructible here,
  since ToolMessage.tool_call_id is required and validated at construction
  rather than the call boundary.
• 073: asserts the vendor body error.type verbatim, where python surfaces the
  exception class name (a contract-permitted style).

Testing

• tests/conformance/test_observability.py: 64 passed, 48 skipped.
• Full tests/: 1456 passed, 414 skipped.
• ruff and pyright clean.

[Copilot]

Pull request overview

Wires tier-1 typed LLM-event observability fixtures (proposal 0057/0058 family) into the YAML conformance harness so the Python implementation is validated against the spec-owned fixtures rather than Python-authored unit-test equivalents.

Changes:

• Moves a set of typed-event fixtures (060–065, 067, 068, 071, 072) into _SUPPORTED_FIXTURES and routes them through the conformance runner.
• Adds new harness machinery: multi-node chain runner, expanded assertion shapes (event_counts, fields_absent_keys, <any-string>), and recursive value matching for structured fields like active_prompt.
• Consolidates shared helpers for mock transport + RuntimeConfig construction and strengthens failure-path assertions (expected_error, call_id invariants).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.