perf(feed): precompute per-mission search index instead of re-normalizing on every keystroke

[guyghost]

Finding (audit point #10, from PR #327's audit table)

apps/extension/src/lib/state/feed.svelte.ts — recomputeFilteredMissions built its search haystack per mission on every search keystroke: it lowercased and space-joined the full mission description (~4 KB) plus title/client/location/source/stack for each mission, on each keystroke. With N missions that's ~N×4 KB of string join + lowercase work per keypress, repeated for every character typed.

Approach

• Extract the haystack construction into a pure helper buildSearchHaystack(mission) — byte-for-byte the same field list, order, non-empty-string filter, ' ' join, and toLowerCase() as the previous inline logic.
• In the store factory, precompute the index once per missions change:

  const searchIndex = $derived(missions.map(buildSearchHaystack));
  const filteredMissions = $deduced... = $derived(recomputeFilteredMissions(missions, searchIndex, searchQuery));

  $derived is lazy and reactive: the index is rebuilt only when missions changes, never on searchQuery changes. Dependency ordering guarantees index/missions positional alignment.
• recomputeFilteredMissions(missions, searchIndex, searchQuery) now only trims/lowercases the query and does searchIndex[index].includes(query) — a plain substring scan over pre-normalized strings.
• Plain string[] on purpose — no reactive collection needed (rebuilt wholesale on missions change), so the svelte/prefer-svelte-reactivity rule isn't triggered and no file-level disable is required.

Before → after per-keystroke cost

	Before	After
Per keystroke	N × (filter + join + lowercase of ~4 KB haystack)	N × substring includes on precomputed strings
Per missions change	—	N × (join + lowercase), once

Memory tradeoff: the index holds ~one lowercased haystack (~size of mission text) per mission in memory. Bounded by the feed size and rebuilt wholesale on missions change — acceptable for a local-first extension feed.

Semantics preservation

Search behavior is identical by construction:

• same fields searched, in the same order: [title, client, description, location, source, ...stack]
• same non-empty-string filter and ' ' join → multi-word queries can still match across field boundaries (joined-haystack semantics preserved)
• same case-insensitive matching (haystack lowercased once instead of per keystroke; query trim().toLowerCase() — equivalent to the previous toLowerCase().trim() since case mapping never produces/removes whitespace)
• substring matching (includes) preserved — no tokenization, whole-word, or diacritic-folding changes
• empty/whitespace-only query still returns missions unfiltered

A dedicated equivalence test (produces identical results to the previous per-keystroke normalization) runs a corpus of 12 representative queries — uppercase, multi-word, cross-field, whitespace-padded, absent terms, empty — against a local replica of the previous inline algorithm and asserts identical results.

House rule: Model → Review → Implement → Verify

This is a behavior-identical performance refactor of the feed search path: no states, events, transitions, or side effects change; filtering decisions are unchanged and remain entirely deterministic model/state logic (no LLM involvement). The authoritative model (src/models/feed-story.model.md) specifies search only at the observable level, which this PR preserves exactly — no model update required.

Tests

Extended tests/unit/state/feed.test.ts with a precomputed search index suite:

• index reused across successive searches on the same missions
• index rebuilt when missions change (query persists across setMissions)
• multi-word queries spanning field boundaries (joined-haystack semantics)
• case-insensitivity for both query and mission fields
• description-only matches
• whitespace-only queries treated as empty; surrounding whitespace trimmed
• equivalence with the previous per-keystroke normalization

Verification

• pnpm --filter @pulse/extension exec vitest run tests/unit/state/feed.test.ts → 18/18 passed
• pnpm --filter @pulse/extension typecheck → clean
• pnpm --filter @pulse/extension lint → 0 errors (8 pre-existing warnings in local-data-reset.contract.ts, untouched)
• Full suite → 4079 passed / 51 skipped, with one pre-existing flaky timeout (local-data-reset-machine.model.test.ts "rehydrates cleanup after adoption…") reproduced identically on the pristine baseline commit — unrelated to this change
• Final push ran the full pre-push gate (ci:check: format + lint + typecheck + test + build) — green
• prettier applied to both touched files

Closes audit point #10.

---

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
pulse	Ready	Preview	Aug 28, 2026 9:38am

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
pulse-dashboard	Skipped		Aug 28, 2026 9:38am

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[Copilot]

🟢 Approval recommended

The refactor is small, behavior-preserving, and is backed by targeted unit tests that assert equivalence with the prior search implementation.

Pull request overview

This PR optimizes the extension feed’s client-side search filtering by precomputing a per-mission normalized search “haystack” when missions change, instead of rebuilding/lowercasing/joining large strings on every keystroke. It fits into the apps/extension/src/lib/state/ state-module layer and preserves observable search semantics while reducing per-keystroke CPU work.

Changes:

• Extracted haystack construction into a pure helper (buildSearchHaystack) and introduced a derived searchIndex that rebuilds only when missions changes.
• Updated filtering to use searchIndex[index].includes(query) with a trimmed/lowercased query.
• Added/expanded unit tests to validate index reuse/rebuild behavior and assert equivalence with the previous inline algorithm across a representative query corpus.

File summaries

File	Description
apps/extension/src/lib/state/feed.svelte.ts	Precomputes a lowercased joined search haystack per mission and filters against a derived index to avoid per-keystroke normalization work.
apps/extension/tests/unit/state/feed.test.ts	Adds a dedicated “precomputed search index” test suite, including an equivalence test against the previous algorithm.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.