fix(lsp): recover from LSP server crashes and stop leaking timed-out requests

[bug-ops]

Summary

• Fixes lsp/client.rs: pending_requests entries leak forever on request timeout #239: LspClient::request never removed its pending_requests entry on timeout, leaking one map entry per timed-out call for the life of the connection. LspClient gains fail_pending_requests to fail stragglers immediately instead of leaving each to time out on its own.
• Fixes mcpls does not recover after its LSP server subprocess dies #249: mcpls previously surfaced a clean error on every tool call once a routed LSP server's child process died, with no recovery until the whole mcpls process was restarted. Translator now detects a dead child process and transparently respawns and re-initializes it before resolving the next tool call for that server.

Details

• Concurrent callers observing the same dead server single-flight on a per-server lock, so only one respawn happens; requests still parked on the old connection are failed immediately via fail_pending_requests.
• A crash-looping server backs off exponentially (1s up to 30s) instead of retrying on every tool call, including the "starts, initializes, then dies again shortly after" loop, not just an outright spawn failure.
• DocumentTracker::forget_server resets per-server document sync state on respawn so the next access re-opens instead of sending a stale didChange; a per-server generation counter closes the race where an in-flight sync against the old connection could otherwise land after a concurrent respawn.
• The respawn path invalidates the crashed connection's cached diagnostics when the crashed server was the diagnostics route for its language, so a crash doesn't leave stale pre-crash diagnostics merged into later polls.
• New Error::ServerUnavailable variant distinguishes "respawn could not proceed" (no config registered, or backing off) from a plain Error::ServerTerminated.
• Known scope trade-off: a respawned server's own push notifications are drained and discarded rather than reconnected to the existing notification pump, so diagnostics push does not resume for it until the whole mcpls process restarts.

Went through several rounds of adversarial review before this PR: initial implementation, three rounds of critique catching a critical stale-diagnostics bug and two significant concurrency races (a document-sync TOCTOU and a backoff-reset bug that let post-init crash loops bypass backoff), and a final code review pass. Also required a non-trivial rebase merge with #260's capability-gating work, since both PRs touched the same client-resolution call sites — respawn-awareness and capability-gating are now composed correctly at every call site (respawn-check-and-recover before capability-gating, since a dead server can't be asked about capabilities).

Test plan

• cargo +nightly fmt --all -- --check
• cargo clippy --all-targets --all-features --workspace -- -D warnings
• cargo nextest run --workspace --all-features --lib --bins (531 passed)
• RUSTDOCFLAGS="-D warnings" cargo doc --no-deps --all-features
• cargo test --doc --workspace --all-features
• New regression tests for both fixes, including a genuine concurrency test proving single-flighted respawn and an integration-level test driving crash-and-respawn through prepare_document
• Capability-gating (LSP server capabilities are negotiated but never checked before dispatching tool requests #240) and respawn tests verified passing together, not just independently

[Copilot]

🟡 Not ready to approve

New regression tests spawn Unix-specific binaries (e.g. sleep, cat, echo) without a Windows-compatible path, which is likely to break the existing Windows CI matrix.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Pull request overview

This PR improves mcpls’s resilience and resource hygiene in the MCP↔LSP bridge by (1) preventing timed-out LSP requests from leaking in-memory bookkeeping and (2) adding transparent, single-flighted LSP server respawn when a routed server subprocess crashes, including reset of per-server document sync state and invalidation of stale diagnostics cache.

Changes:

• Fix LspClient::request timeout cleanup by removing timed-out entries from pending_requests, and add fail_pending_requests to immediately unblock stragglers when a connection is discarded.
• Add crash detection (LspServer::has_exited) and Translator respawn supervision with per-server single-flight + exponential backoff, plus document sync reset (DocumentTracker::forget_server) and diagnostics cache invalidation on respawn (diagnostics-route only).
• Wire respawn configuration registration through startup (register_servers), add Error::ServerUnavailable, and document behavior in CHANGELOG.md.

File summaries

File	Description
crates/mcpls-core/src/lsp/lifecycle.rs	Stores the child handle and adds has_exited(); adds a regression test for crash detection.
crates/mcpls-core/src/lsp/client.rs	Fixes timeout leak in pending_requests and adds fail_pending_requests, plus regression tests.
crates/mcpls-core/src/lib.rs	Passes per-server spawn configs into the translator and wires in a shared diagnostics cache handle.
crates/mcpls-core/src/error.rs	Adds Error::ServerUnavailable to distinguish “dead” vs “could not respawn/backing off”.
crates/mcpls-core/src/bridge/translator.rs	Implements respawn detection/recovery, single-flight locking, backoff, and cache invalidation integration.
crates/mcpls-core/src/bridge/state.rs	Adds per-server generation tracking and forget_server to prevent post-respawn sync races.
CHANGELOG.md	Documents respawn-on-crash and timeout-leak fix under Unreleased entries.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 2
• Review effort level: Lite

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.