ci: add MCP SDK compatibility workflow

[Oaklight]

Summary

• Add .github/workflows/mcp-compat.yml — a dedicated CI workflow that tests MCP integration against both v1.x and v2.x SDK lines, with weekly schedule (Wednesday 04:41 UTC) and manual dispatch with optional version override
• Follows the rosetta-compat.yml pattern: matrix testing, per-leg artifact upload, auto-issue filing with mcp-compat label on failure
• Remove flaky wall-clock elapsed < 4.0 assertions from 9 timeout tests — the ErrorResult + "timed out" checks already verify correctness without depending on execution speed

Test plan

• All 9 timeout tests pass locally after removing elapsed assertions
• Full MCP test suite (113 tests) passes locally
• mcp-compat GitHub label created
• CI passes on this PR
• Manual workflow dispatch works: gh workflow run mcp-compat.yml

[elena-oaklight]

LGTM 👍

Workflow (mcp-compat.yml):

• Matrix strategy with fail-fast: false ensures both SDK lines get tested independently
• Version resolution logic is solid — validates format, correctly skips mismatched matrix legs on manual dispatch
• Issue deduplication via v${line}.x title match prevents spam; comment-on-existing is a nice touch
• The v1 extra deps (httpx>=0.28.1, websockets) handling looks correct

Test cleanup:

• Good call removing the elapsed < 4.0 assertions — wall-clock checks are inherently flaky under CI load
• The remaining ErrorResult + "timed out" checks are sufficient to verify timeout correctness

One minor note: the cron schedule 41 4 * * 3 (Wed 04:41 UTC) is fine, though if you want separation from rosetta-compat, maybe stagger by a few hours. Not blocking.

[clementine-oaklight]

Clean PR — workflow follows the rosetta-compat.yml pattern correctly, and the timeout test cleanup removes the right assertions.

Workflow (mcp-compat.yml):

• Cron timing, matrix with fail-fast: false, version resolution with validation, artifact upload, and auto-issue deduplication all look solid.
• httpx>=0.28.1 + websockets for v1 line only makes sense if v2 bundles these differently.

Test changes:

• Removing elapsed < 4.0 is correct. Wall-clock assertions are inherently flaky under CI load; the ErrorResult + "timed out" checks already prove correctness.

Minor flags:

1. Issue title collision risk (L118-119) — dedup matches v${line}.x substring in title. If someone manually edits the title to remove that substring, a second issue gets created on the next failure. Could add a hidden HTML comment marker in the body and match that instead, but this is an edge case.

2. Version resolution when v2 not yet published — mcp>=2.0.0,<3 would fail if the v2 line doesn't exist on PyPI yet. If this is expected (v2 exists), no issue. If v2 is future-facing, the step should tolerate "no matching distribution" gracefully.

Neither is blocking — approve once CI is green.