fix(mcp): give tools/call its own request budget and stop undercutting long executions

[asto18089]

Summary

MCP tool calls were killed by two stacked short budgets: crates/mcp's generic REQUEST_TIMEOUT (120s) applied to all requests including tools/call, and the TUI pool's default_execute_timeout (60s). A legitimate tool call that runs minutes — builds, test suites, scrapes, remote jobs — was killed mid-flight, and the model's retry compounded the cost. The read-wait also undercut both: recv bounded every response wait by the 120s read knob, so even a raised execute_timeout fired first at the inner read and marked the connection Disconnected.

Changes:

• crates/mcp: tools/call gets a dedicated CALL_TOOL_TIMEOUT (30 min, matching the TUI pool default); every other request (initialize, lists, resources) keeps the 120s generic budget.
• TUI pool: default_execute_timeout 60s → 30 min; per-server/global execute_timeout overrides still win via the existing effective_execute_timeout seam.
• Read-wait: the inner read budget for a request is now max(read_timeout, request budget), so the read knob can no longer silently defeat a raised execute budget; requests whose budget doesn't exceed the read knob keep failing fast at it.
• HTTP: the client ceiling is built from max(read, execute) so HTTP servers get the same semantics.

Known trade-offs, disclosed for review:

• prompts/get routes through effective_execute_timeout, so it inherits the new default (comment in code notes this; the override knob is the intended way to keep it tight).
• All MCP calls serialize on one pool mutex held for the call, so a genuinely wedged tools/call now blocks sibling-server calls/status for up to 30 min (previously ≤60s) unless the turn is interrupted — same shape as before, longer tail; the per-server execute_timeout override is the escape hatch. Happy to explore hold-lock-only-for-dispatch as a follow-up.

Coordination: upstream #6711 reworks engine stream-open budgets — different surface, no file overlap.

Adapted from the Pinvou fork's timeout audit (Pinvou/CodeWhale d349f2537).

Testing

• cargo fmt --all -- --check
• cargo clippy -p codewhale-mcp -p codewhale-tui --all-features --locked (clean under the CI allow list)
• cargo test -p codewhale-mcp --lib — 83 passed; cargo test -p codewhale-tui --lib mcp:: — 251 passed. New pins: a slow tools/call outlives an injected generic budget while resources/read still fails at it; a tool response after the read knob still completes within the execute budget (red without the widening); requests under the read knob still fail at it; the HTTP ceiling covers the execute budget

Checklist

• Updated docs or comments as needed (docs/MCP.md + zh minimal example now shows the new default)
• Added or updated tests where relevant
• No CHANGELOG.md changes