fix(agent): let context length resolution fall through when failed to get context_length from provider

[nocoo]

What does this PR do?

Fixes an early return in get_model_context_length() that causes custom endpoints without context_length in their /v1/models response to default to 128K, even when later resolution steps (hardcoded defaults, models.dev, OpenRouter) could correctly resolve the context length.

Example: A proxy serving claude-opus-4.6-1m that omits context_length from its model metadata would return 128K instead of 1M, because the hardcoded DEFAULT_CONTEXT_LENGTHS entry "claude-opus-4.6" (a valid substring match at step 8) was never reached.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/model_metadata.py: Remove the early return DEFAULT_FALLBACK_CONTEXT at the end of the custom endpoint probe block (step 2/3). The function now falls through to remaining resolution steps (4–8) when the endpoint probe does not yield a context_length. The logger.info message is preserved for observability.
• tests/agent/test_model_metadata.py: Update test_custom_endpoint_without_metadata_falls_through_to_hardcoded_default (formerly ..._skips_name_based_default) to assert the new fall-through behavior. Add two new regression tests:
  • test_custom_endpoint_no_context_length_falls_through_to_defaults — endpoint returns model metadata without context_length
  • test_custom_endpoint_no_match_falls_through_to_defaults — endpoint returns unrelated models, requested model falls through to hardcoded defaults

How to Test

from agent.model_metadata import get_model_context_length

# Before fix: returns 128000
# After fix: returns 1000000
result = get_model_context_length(
    "claude-opus-4.6-1m",
    base_url="http://localhost:7024/v1",  # proxy without context_length in /models
    api_key="test-key",
)
assert result == 1000000

python -m pytest tests/agent/test_model_metadata.py tests/agent/test_model_metadata_local_ctx.py -q

Checklist

Code

• I have read the contributing guidelines
• Commit messages follow Conventional Commits
• I checked for duplicate PRs
• Changes are focused and minimal
• pytest tests/ -q passes
• Tests added for the fix
• Tested on macOS

[alt-glitch]

Related to #15563, #12059 — same early-return bug in get_model_context_length() step 2 short-circuiting to 128K.

[alt-glitch]

Related to #15563, #12059 — same early-return bug in get_model_context_length() step 2 short-circuiting to 128K.

[teknium1]

Thanks for isolating the custom-endpoint early-return issue.

Problems

• Current main already fixes the catalog-match case through the targeted fallback at agent/model_metadata.py:2238-2258, with regression coverage at tests/agent/test_model_metadata.py:999-1053 (commit 2e61de063).
• This PR's broader fall-through changes precedence beyond that fix, but its added tests only exercise hardcoded catalog matches. They do not demonstrate a result obtained specifically from a later provider, models.dev, or OpenRouter lookup.

Suggested changes

• Add a regression whose expected result cannot be produced by DEFAULT_CONTEXT_LENGTHS, and document why the later metadata source is authoritative for an incomplete custom endpoint.
• Clarify the intended source precedence before extending the current bounded catalog fallback.

This is an automated hermes-sweeper review.