Local suite deadlines are too tight for dist entrypoint and real-git worktree tests

[willwashburn]

Observed

During the full local validation of PR #207 at 62684be10da10de6900f8bc37615a37e9cfe373c, five tests hit Vitest's fixed 5-second ceiling:

• src/__tests__/dist-entrypoints.test.ts — published dist entrypoints > are importable by Node ESM consumers
• four cases in src/git/agent-worktree.test.ts, which exercise sequential real git operations in temporary repositories/worktrees

The dist import completed in 5.04 seconds when rerun with a 20-second ceiling. The worktree file passed 9/9 under the same wider ceiling. Those diagnostics establish deadline sensitivity; they do not prove an external cause or that the behavior is pre-existing.

Exact-head Actions run 30553216460 passed the full package suite, so these failures do not gate PR #207. They remain defective local-test ergonomics: a busy machine can fail an otherwise-correct suite on a magic wall-clock threshold. This is the same class as #208.

Required outcome

• Replace the incidental fixed 5-second ceilings with bounds proportional to the operations being exercised, or assert the behavioral invariant without including unrelated process scheduling and filesystem setup in one magic deadline.
• Preserve a real upper bound so an actually hung Node import or git process still fails.
• Verify the affected tests both in isolation and as part of the loaded full local suite.
• Record enough per-test timing evidence to distinguish deadline sensitivity from a functional regression.

Scope

Test infrastructure only. Do not weaken the production timeout or process-cleanup behavior being exercised.