feat(search): collapse over-long separators in recalled snippets

[niemst]

What / Why / How

ctx_search echoes indexed content verbatim, so decorative separators (=×80 banners, long --- rules) are recalled in full — spending context tokens for ~zero information. extractSnippet now collapses any single non-whitespace char repeated 12+ times to ccc…×N, lossless about the original length.

• Threshold 12 leaves real markup intact: --- rules, ***/___, ~~~ fences, @@ diff hunks.
• Applied only to the returned snippet, after FTS5 highlight offsets are computed — highlight positions are unaffected.
• \S-only, so whitespace/newline structure is never touched.

Before

--- [session | … | file:/tmp/report.txt] ---
================================================================================
flow_id=306058 …

After

--- [session | … | file:/tmp/report.txt] ---
===…×80
flow_id=306058 …

Affected platforms

• All platforms

Recall snippet formatting is platform-agnostic.

Test plan

Added 5 cases to tests/core/search.test.ts (§ extractSnippet collapses over-long separators): 80-char banner → ===…×80, short markup untouched, lossless ×N, whitespace runs untouched, collapse survives the windowed/truncated path. npm test: 4244 pass (2 pre-existing macOS locale/PWD env failures, unrelated to this change). npm run typecheck: clean.

Bundles not rebuilt on purpose — bundle.yml regenerates them on merge to main, and next's committed bundle is already stale vs source, so a local rebuild would only add unrelated churn.

Checklist

• Tests added/updated (TDD: red → green)
• npm test passes (2 unrelated pre-existing env failures)
• npm run typecheck passes
• Docs updated if needed (n/a — internal formatter)
• No Windows path regressions (no path code touched)
• Targets next branch