fix(copilot): tolerate SDK metadata parsing errors when listing models

[jrob5756]

Summary

Workflows that configure reasoning.effort (or workflow-wide
runtime.default_reasoning_effort) on the Copilot provider were failing with:

Dialog turn failed: Missing required field 'multiplier' in ModelBilling

after exhausting the retry loop. The bug affects every named Copilot model
(see "Blast radius" below) — it just wasn't caught earlier because few
workflows exercise the reasoning-effort path on Copilot.

Root cause

github-copilot-sdk 0.3.0 eagerly parses every entry in its
models.list response with dataclass from_dict helpers. The
ModelBilling.from_dict parser
(client.py:496-507)
treats multiplier as required and raises
ValueError("Missing required field 'multiplier' in ModelBilling") if it's
absent — and none of the currently-served models include it. Because the
parsing is done in a list-comprehension, a single malformed entry kills the
whole list_models() call, and the SDK's response cache is only populated
after a successful parse, so every call from a fresh session fails the same way.

Both _validate_reasoning_effort_for_model (called before every
create_session when reasoning.effort is set) and get_max_prompt_tokens
caught only (TimeoutError, ProviderError, OSError, RuntimeError) at the
SDK boundary, so the ValueError leaked through. In the reasoning-effort
path it then poisoned the retry loop and surfaced as the user-visible
Dialog turn failed: ... error. (get_max_prompt_tokens was rescued by
the engine's outer except Exception at
engine/workflow.py:1071,
so context-window metadata was silently unavailable rather than fatal.)

Blast radius

Inspected the live response on this machine — 18 of 19 returned models ship
a billing object without multiplier:

claude-sonnet-4.6, claude-sonnet-4.5, claude-haiku-4.5,
claude-opus-4.7, claude-opus-4.7-1m-internal, claude-opus-4.7-high,
claude-opus-4.7-xhigh, claude-opus-4.6, claude-opus-4.6-1m,
claude-opus-4.5, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex,
gpt-5.2-codex, gpt-5.2, gpt-5-mini, gpt-4.1

The only "clean" entry is auto, which has billing: None so the parser
short-circuits.

Fix

Broaden both catches to Exception (still excludes
asyncio.CancelledError / KeyboardInterrupt / SystemExit, all of
which are BaseException subclasses) and document the SDK schema-failure
case at both call sites. This matches the existing docstring contract:
"capability metadata must never block a workflow that the SDK might
otherwise accept."

The narrow tuple was written before anyone realized the SDK would raise
non-OSError exceptions in normal operation.

Tests

• Replaced test_unexpected_exception_propagates (which asserted the
  opposite of the docstring contract) with
  test_value_error_from_sdk_parser_returns_none, using the exact
  upstream error message for clarity.
• Added test_value_error_from_sdk_parser_skips_validation covering the
  reasoning-effort path: confirms the workflow proceeds normally and
  that reasoning_effort is still forwarded to create_session.
• Full suite green: 2579 passed, 15 skipped, 29 deselected in 7m36s.

Note

This is an internal-model regression on the Copilot SDK side, not worth
filing upstream — the missing field is a temporary backend quirk and the
SDK will likely make multiplier optional in a future release. Our fix
hardens the code path regardless.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com