fix(code-intel): order transport failure publication

[Qiyuanqiii]

Summary

• publish the terminal transport signal before releasing pending LSP requests with the same failure
• prevent a diagnostic transport crash from racing into the push-diagnostics fallback as a successful empty result
• add a deterministic ordering regression test for the transport lifecycle contract

Root cause

transport.setFailure marked the transport closed, but then delivered errors to pending request channels before closing transport.done. A released request could resume in that interval. queryDiagnostics uses done to distinguish a terminal transport failure from an ordinary pull-diagnostics RPC error, so the process-exit error could be misclassified and converted into a successful push fallback.

The test helper deliberately exits during textDocument/diagnostic. Under the race detector the affected request occasionally returned nil, leaving the crash masked. The same unchanged code alternated between pass and failure on the #159 head and its merge commit.

This change closes done before publishing pending request errors. Every released waiter therefore observes the terminal state. The new unbuffered-channel test makes that ordering contract deterministic and fails against the previous implementation.

This is a pre-existing Code Intel lifecycle race introduced with #119. It is unrelated to the CSS-only changes in #158 and to the Command Runtime changes in #159.

Validation

• reproduced before the fix: go test -race -count=100 -timeout 10m ./internal/codeintel -run '^TestDiagnosticsDoesNotMaskTransportCrashAsPushFallback$'
• fixed scenario, 1000 consecutive race-detector runs
• complete Code Intel race/fault suite: go test -race -count=1 -timeout 5m ./internal/codeintel
• complete Code Intel package: go test -count=1 -timeout 5m ./internal/codeintel
• go vet ./internal/codeintel
• git diff --check

Repairs the failure observed on main@b42eccd in CI run 32749900859.