[alex09x]

Problem

Issue #1594 describes Rust code_references lookups for common unqualified names taking up to about 20 seconds in multi-target discovery. A batch of several symbols also needs to preserve time for later entries and return useful partial results when one reference scan exhausts its share.

Change

• Give each unqualified code_references batch symbol a default 2 second share when no path hint or explicit budget is supplied.
• Pass the share into execute_tool without an outer timeout that would abort candidate resolution and discard the structured partial result. The batch stops starting later symbols when its shared budget is exhausted.
• Do not apply the default unqualified cap when the caller supplies a path/file hint; explicit budgets remain in control.
• Make the multi-target reference scan use the remaining budget_ms from the start of handle_references, cap it at the existing 8 second maximum, and return an explicit partial-results note when the scan reaches its deadline.

Reproduce and verify

Unit regressions verify that an unqualified name receives the 2 second budget, a path-scoped name retains the shared budget, an explicit budget overrides the default cap, and a partial result returned just after the 2 second share survives the batch wrapper. The existing multi-candidate test verifies that execute_tool returns candidate/partial guidance when its own shared budget expires. Reference-scan tests verify that shorter budgets are honored, long budgets retain the 8 second cap, and an expired deadline returns partial guidance without issuing an LSP query.

Measurements

No performance claim. The batch regression waits 2.1 seconds to prove a returned partial result is not discarded at the former outer 2 second timeout.

Checks

Revision 3c369d284cd72a7b42102d8820469ac49350bcc5, Linux build node:

• cargo fmt --all -- --check: exit 0.
• cargo clippy -p prod-code-mcp --all-targets: exit 0 with existing warnings (125 warnings in the test build); this is not a strict-clean lint claim.
• cargo test -p prod-code-mcp --lib unqualified_reference_batch_ -- --nocapture: 2 passed, 0 failed.
• cargo test -p prod-code-mcp --lib multi_target_deadline -- --nocapture: 2 passed, 0 failed.
• cargo test -p prod-code-mcp --lib multi_candidate_shared_budget_returns_partial_result_when_expired -- --nocapture: 1 passed, 0 failed.
• prod-code impact --base origin/main --depth 1 --json: exit 0; it selected the direct batch/reference tests and reported incomplete caller coverage at depth 1. No broad suite was run.

The current head also integrates #1715. Its merge commit adds the definition-body worktree fix, tests and changelog only; batch/reference sources and tests are unchanged from f2c94f47.

Closes #1594

[alex09x]

Independent read-only P0–P3 review of 103b521: APPROVE, no findings. The batch/reference deadlines and partial-result regressions passed on the current base.

[alex09x]

Independent read-only P0–P3 re-review of 3c369d2: APPROVE, no findings. The current-base merge contains no changes to the batch/reference implementation or tests; all focused checks are recorded in the PR.