fix(bridge): signal degraded positions and evict documents on the LRU cap

[bug-ops]

Summary

Two independent fixes in the bridge subsystem, both surfaced by continuous-improvement cycle 035 against v0.5.0 (commit bfe82b0).

• Position-encoding conversion now sets a positions_degraded flag on every DTO whose fields come from position conversion (hover, definition, references, locations, rename, format, code actions, document/workspace symbols, call hierarchy, inlay hints, diagnostics) whenever a column could not be converted for encoding — disk-read budget exhausted, unresolvable server path, oversized file, invalid UTF-8, or a line past EOF. Previously the only signal was a tracing::warn! in mcpls's own log; a caller had no way to tell that the tail of a result set might carry silently wrong columns.
• DocumentTracker now evicts the least-recently-used unlocked document (sending textDocument/didClose to the LSP server) when open() would exceed workspace.max_documents, instead of permanently refusing every new document for the rest of the process lifetime once the cap is reached. Eviction skips documents with an in-flight operation (tracked via path_locks) or unsaved local edits (disk().is_some() gate), falling back to the existing DocumentLimitExceeded error only when no evictable candidate remains.

Details

positions_degraded (#497) — the flag is set directly at the "passing server column through unconverted" decision point in to_lsp/to_mcp (not derived from the disk-read-budget bookkeeping, which only covers one of several causes), so it fires on the first degraded lookup in a response, not just after the budget is fully exhausted. Translator::merge_diagnostics ORs the flag across the pull and cached diagnostics contexts. All affected MCP tool descriptions and DTO doc comments were updated to describe the flag; tool_surface.json was regenerated accordingly.

Document eviction (#495) — DocumentTracker has no LspClient access, so evictions are queued as EvictedDocument { path, uri, synced_servers } and drained by Translator::notify_evicted_documents, which sends textDocument/didClose after every ensure_open call (including on that call's own error path, so a queued eviction from a prior lookup is never lost). DocumentState::last_accessed tracks LRU order, bumped on every ensure_open/update against a tracked document.

A narrow, non-default-config race is documented but deliberately not structurally closed: PathLockGuard covers ensure_open/update, not the LSP round-trip that follows, so at max_documents: 1-2 with concurrent calls, two in-flight documents could in principle evict each other. This is gated behind a config far below the default (100), and closing it structurally would require widening lock scope across every MCP handler — out of scope for this PR. A follow-up issue will track it.

Test plan

• cargo +nightly fmt --all -- --check
• cargo clippy --all-targets --all-features --workspace -- -D warnings
• cargo nextest run --workspace --all-features --lib --bins (1016 passed, 1 skipped)
• RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace
• New regression tests: positions_degraded flips true end-to-end through a full mocked handler call and through the merge_diagnostics OR-merge; DocumentTracker LRU-vs-insertion-order distinction, lock-skip fallback to DocumentLimitExceeded, unsaved-edit eviction guard, and wire-level textDocument/didClose on eviction

Closes #497
Closes #495