fix(search): find whole-word matches past the per-file substring cap

[DevMello]

Summary

Keyword search with whole_word could return nothing while a real match sat in
the file. --max-count caps ripgrep's unfiltered hits, but whole-token filtering
runs app-side, so a file whose first 50 hits are all substring-only loses every
real match below that point and still reports no truncation.

What changed

• Pass --max-count only on the substring path, where the whole-token filter is
  a no-op.
• Apply the per-file cap after filtering when whole_word is set, like
  searchDerivedMarkdown already does. Non-whole-word searches are unchanged.

Validation

• pnpm typecheck
• Focused tests for the affected behavior
• Renderer build or E2E coverage, when the change affects the UI

Two new cases in server/keyword-search.test.ts, red before and green after.

UI and visual baselines

• This PR changes a rendered UI surface or visual state.
• This fork PR allows maintainer edits if a reviewed Linux visual baseline update is needed.
• This PR does not change a visual surface, or existing visual baselines remain valid.

Documentation

• Updated the relevant design-docs/ and code-review/ contract, when this changes documented behavior or invariants.
• No documentation update is needed.