[kalvinnchau]

Problem

A WebKit timeline recording of cmd+k channel navigation with many agents running shows the search palette's SearchResults render taking about 35% of JS samples. About 2,800 of 11,390 samples (~25% of all JS) are in archiveHides(), which calls snapshot().archived.includes(pubkey) for every person and agent candidate on every render. Each render therefore costs candidates × archived identities, and agent traffic re-renders the open palette repeatedly.

Change

archiveHides builds one Set per archived list, cached in a WeakMap keyed by that list. The store freezes a new list for each snapshot, so a cached Set always matches its snapshot. One agent-choices test fixture changed its archived list in place; it now publishes a new frozen list on each read, as the store does.

No behavior changes

Same answers for every input: the viewer is never hidden, missing archives hide nobody, and every other key is hidden exactly when it is in the snapshot's archived list.

Evidence

Both implementations as they are in the source, interleaved, one call per candidate per render. Median ms per render, 6 runs per cell, on 2 Linux and 2 macOS workstations (macOS also in JavaScriptCore, the desktop engine).

archived × candidates	Linux node: main	Linux node: this PR	Linux Δ %	macOS JavaScriptCore: main	macOS JavaScriptCore: this PR	macOS Δ %
100 × 500	0.50 ms	0.02 ms	−96.3%	0.38 ms	0.007 ms	−98.0%
1,000 × 1,000	9.32 ms	0.04 ms	−99.6%	8.61 ms	0.015 ms	−99.8%
1,000 × 3,000	26.86 ms	0.10 ms	−99.6%	25.43 ms	0.044 ms	−99.8%
5,000 × 3,000	80.83 ms	0.11 ms	−99.9%	71.95 ms	0.064 ms	−99.9%

Lower is better. Each value is JS time for one palette render's archive checks.
macOS node (V8) gives the same shape: 62.38 → 0.07 ms (−99.9%) at 5,000 × 3,000.

Checks: tsc and biome are clean. The identity-archives, SearchResults, agent choices, mention admission, session agent admission, ChannelMembersDialog and solo suites pass (494 tests). There is no new test, because only speed changes.

🤖 Generated with Claude Code

[wesbillman]

🤖 Star Lord automated source review via Wes’s account.

No actionable findings in the diff from base fae935a363bd682d128ba7962fd237b99062902b to head 2f87dc5d9851e0dee7d7c1cf9426c0b628d3604b.

Reviewed the per-array WeakMap cache, archive snapshot replacement and clearing, retained visibility during loading/errors, viewer exemption, and search/chooser/mention callers. The current producers preserve each array’s contents, and the changed fixture publishes a fresh copy on refresh. This supports reusing a Set for each list without changing membership answers.

Coverage: Rocket (systems/performance/lifecycle), Mantis (correctness/security), Groot (product/callers), and Star Lord (integration and public description/commit inspection). Source-only: no tests, app workflows, browser/native interaction, or benchmarks were run. The PR’s timing and test results were not independently verified.