perf(search): fold people names once across palette keystrokes - #2570

Co-authored-by: BuzzPRReviewAssistant1 <a581a7e8115222acdc32ffb14bfce55d48c24478d98e155f8eba126b69b83c7b@buzz.block.builderlab.xyz>

* origin/main: (71 commits)
  test(activity): scope preview assertions to the request row (#2843)
  perf(search): fold people names once across palette keystrokes (#2570)
  fix(github): keep PR watches running when localStorage is full (#2699)
  Make drafts and sent part of Messages (#2728)
  fix(canvas): render every Canvas within the save limit (#2839)
  test(messages): wait for the row measurement frame before thread catch-up dwell (#2837)
  Keep agent coordination history readable after work finishes (#2454)
  test(mentions): serve the mentions fixture as a build so cold loads fit (#2776)
  test(browser): let a late publication land after its page closed (#2565)
  test(relay): wait for the verified owner before a guest edits an agent (#2413)
  test: wait for the companion default width measurement (#2750)
  test(browser): preserve conversation scroll while revealing actions (#2715)
  test(browser): wait for link hit-testing after thread close (#2792)
  perf(messages): skip micromark for plain prose bodies (#2780)
  perf(inbox): load the emoji name table on first use (#2774)
  perf(messages): cut main-thread work while a busy channel is open (#2729)
  Refine agent Attention cards, watch details, and status controls (#2717)
  fix(bestie): remove a Bestie from its agent card on any device (#2720)
  feat(agents): preview Builderlab agent card links with Add to Buzz (#2719)
  Restore macOS Cmd-W window closing when no tab handles it (#2443)
  ...

Signed-off-by: Ravneet Arora <[email protected]>

…2726)

This PR combines two changes that cut main-thread work in large
channels. The conversation header no longer re-renders its member
avatars and members button on traffic in other channels, and matching
people by name folds each name once instead of on every keystroke. Both
costs grow with the member count.

## Problem

**Header.** #2626 moved `ChannelMemberAvatars` into every conversation
header through `ChannelHeaderIdentity`. It also gave
`ChannelMembersButton` its own subscription to the whole channel list.
Both components re-render on every list publish, which means every live
message in any channel.

On each of those renders, both components run a `list.channels.find`
over every channel. The avatars also join, filter and sort every member
of the open channel. With a 5,080-member channel open, other channels'
traffic keeps re-rendering a header whose content has not changed.

**Name matching.** Typing in the Manage members filter of a large
channel re-scans every member on every keystroke. #2356 raised the
per-member cost: the filter now checks the verified display name,
username and full name, which is about 3 `matchPerson` calls per member
where v0.7.10 made 1. #2585 added a similar scan of every cached profile
to the New message recipient picker.

For each name it checks, `matchPerson` folds every code point through
`normalize("NFKD")`, an accent regex and `toLowerCase`, and folds the
query again on every call. The filter also bech32-encodes each member's
key on every keystroke to match npubs. In a 5,080-member channel this
adds up to seconds of main-thread work while the user types a short
query.

## Change

**Header.**
- `useHeaderChannel` is a new helper in `listed-channel.ts`. It is
`useListedChannel` plus the existing `channels.get` fallback for
channels the list does not hold. Each header component selects only what
it renders through it: members, the Solo flag, or the access summary.
All of these are referentially stable, so unrelated list publishes skip
the render.
- `memberAccess` returns one of four shared frozen objects instead of a
fresh object.
- `ChannelMemberAvatars` memoizes the member key, the agent map and the
preview.

**Name matching.**
- `foldPoint` handles ASCII with a plain `toLowerCase`. ASCII has no
accents, decompositions or final sigma, so the fold is the same.
- `ChannelMembersDialog` folds the query once per render and matches
each name with `matchTier`, which reuses the folded-name cache #2570
added. It also keeps one npub map for as long as it is open, so each
member's npub is encoded at most once.
- Rebased onto #2570: the earlier per-dialog `folds` map in
`matchPerson` was dropped, because #2570's cache already keeps each name
folded across keystrokes.

## No behavior changes

**Header.** The lookup order is the same as before: the list entry
first, then `channels.get`. A membership change still produces a new
`members` array, so the header updates. Previews and activity reuse the
channel summary's `members`, so they no longer re-render the header.

New tests cover three cases:
- An unrelated list publish commits nothing.
- A membership change updates both the button and the avatars.
- A channel missing from the list renders through the fallback, and
still skips renders on unrelated publishes.

**Name matching.** The ASCII path gives the same fold as before.
`matchTier` gives the same tier as `matchPerson` (the #2570 test checks
this), and an empty query still matches at tier 0.

The npub map is dropped when the dialog closes and holds one entry per
member.

## Evidence

The table compares main (`8c0cb0c1e`) with this PR as a whole. All runs
are interleaved, with the same two builds on both sides.
- **Browser:** production Chromium at 4× CPU throttling with 128 DMs,
1,001 participants, read state and a 5,080-member channel, while another
channel receives a message every 200 ms. **Idle** is 3 s with no input
while the 5k channel is open. **Switch** opens Cmd+K, types the channel
name and presses Enter. **Members dialog** opens Manage members for the
5k channel and types a 6-character query one key at a time; 1,000
members have profiles. Two Linux and two macOS workstations, 36 samples
per cell; values are the mean of the per-workstation medians.
- **Typing latency:** the same app with 5,080 cached member profiles
that share name prefixes, a deliberate worst case. Each value is the
median Event Timing duration from keydown to the next paint, over 6
interleaved pairs on one Linux and one macOS workstation. Event Timing
quantizes to 8 ms and only reports events of 16 ms or more. **Mention**
types `@Member004` in the composer, **palette miss** types `zzzzzz` in
Cmd+K, and **prose** types `ordinarytext`. Each key waits for the next
rendering opportunity before the following key.
- **Header micro-benchmark:** renders the header with 400 channels and
5,080 members, then delivers 50 live messages to another channel.
- **Picker micro-benchmark:** runs the New message recipient picker's
profile scan, typing 6 keys.
- Micro-benchmark cells are the mean of two workstations per OS, each
the median of 6 runs.

| metric | Linux: main | Linux: this PR | Linux Δ % | macOS: main |
macOS: this PR | macOS Δ % |
|---|---|---|---|---|---|---|
| idle 5k, script | 226.0 ms | 182.5 ms | −19.2% | 150.8 ms | 126.7 ms |
−15.9% |
| idle 5k, task | 367.5 ms | 323.1 ms | −12.1% | 266.6 ms | 240.5 ms |
−9.8% |
| switch 5k, script | 580.5 ms | 563.1 ms | −3.0% | 325.5 ms | 312.9 ms
| −3.9% |
| switch 5k, blocking | 279.1 ms | 264.3 ms | −5.3% | 116.8 ms | 115.6
ms | −1.0% (noise) |
| members dialog, script | 3,272.1 ms | 1,834.2 ms | −43.9% | 2,643.0 ms
| 1,173.2 ms | −55.6% |
| members dialog, task | 3,455.5 ms | 2,055.1 ms | −40.5% | 2,863.3 ms |
1,321.1 ms | −53.9% |
| members dialog, blocking | 2,764.1 ms | 1,327.0 ms | −52.0% | 2,039.2
ms | 650.7 ms | −68.1% |
| mention typing, keydown to paint | 448 ms | 200 ms | −55.4% | 272 ms |
112 ms | −58.8% |
| palette miss, keydown to paint | 136 ms | 104 ms | −23.5% | 88 ms | 72
ms | −18.2% |
| prose typing, keydown to paint | 24 ms | 24 ms | 0% | 16 ms | 16 ms |
0% |
| header renders for 50 messages | 50 | 0 | −100% | 50 | 0 | −100% |
| header time for 50 messages | 58.2 ms | 1.5 ms | −97.4% | 32.8 ms |
1.1 ms | −96.6% |
| picker scan, 1,000 profiles | 42.4 ms | 17.3 ms | −59.2% | 24.3 ms |
9.8 ms | −59.5% |
| picker scan, 5,000 profiles | 213.3 ms | 86.8 ms | −59.3% | 119.1 ms |
48.8 ms | −59.0% |
| picker scan, 10,000 profiles | 431.0 ms | 178.0 ms | −58.7% | 240.5 ms
| 98.9 ms | −58.9% |

Lower is better. These numbers were measured with the earlier per-dialog
`folds` map, before the rebase onto #2570. Blocking is
long-animation-frame blocking time. Typing rows have 162 mention keys,
108 palette keys and 216 prose keys per build; one Linux palette key of
this PR had no Event Timing entry, so that cell covers 107 keys. The
browser harness runs the built frontend with modeled HTTP, not the
native app.

**Larger rosters.** A standalone Node bench of the whole member
predicate over a typing session takes 0.22–0.40× of the old time. It
covers 5k, 9k, 20k and 50k members, with and without names. The first
keystroke after the dialog opens pays the one-time fold and encode: on
large rosters without names it is up to about 1.2× the old cost, and
later keystrokes are 10× faster or more.

Checks:
- The channels, search, mentions, direct-messages and channel-members
unit suites pass.
- `tsc` and `biome` pass.
- After the rebase: the search and channels suites pass (1,163 tests).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Signed-off-by: Kalvin Chau <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>

* origin/main:
  fix(sidebar): rename Remove from Team to Close conversation (#2845)
  perf(channels): cut header and name-matching work in large channels (#2726)
  fix(browser): silence a browser page once its tab goes away (#2721)
  Community Mesh: shared compute plugin (including agent type) (#941)
  perf(relay): cut main-thread work from profile updates (#2730)
  fix(composer): lead the channel placeholder with its sidebar icon (#2829)
  test(activity): scope preview assertions to the request row (#2843)
  perf(search): fold people names once across palette keystrokes (#2570)
  fix(github): keep PR watches running when localStorage is full (#2699)
  Make drafts and sent part of Messages (#2728)
  fix(canvas): render every Canvas within the save limit (#2839)

Co-authored-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Co-authored-by: Tyler Longwell <[email protected]>
Signed-off-by: Tyler Longwell <[email protected]>

…leanup

* origin/main:
  feat(agents): let app-hosted agents read their owner's channels and members (#2745)
  NEXUS-4271: Expose installed plugin status (#2754)
  fix(menu): stop clipping lifecycle actions behind a hidden scrollbar (#2716)
  feat(agents): give approved external apps the selected agent's full Buzz tools (#2707)
  AIDA-2549: Keep Claude settings drafts in sync (#2841)
  fix(channels): keep a channel's side panels open when returning to it (#2838)
  fix(builderlab): clarify remote-agent account availability (#2764)
  fix(sidebar): rename Remove from Team to Close conversation (#2845)
  perf(channels): cut header and name-matching work in large channels (#2726)
  fix(browser): silence a browser page once its tab goes away (#2721)
  Community Mesh: shared compute plugin (including agent type) (#941)
  perf(relay): cut main-thread work from profile updates (#2730)
  fix(composer): lead the channel placeholder with its sidebar icon (#2829)
  test(activity): scope preview assertions to the request row (#2843)
  perf(search): fold people names once across palette keystrokes (#2570)
  fix(github): keep PR watches running when localStorage is full (#2699)
  Make drafts and sent part of Messages (#2728)
  fix(canvas): render every Canvas within the save limit (#2839)

# Conflicts:
#	.github/workflows/release.yml

…ime-hint

* upstream/main:
  ci: give the native job room for a cold Rust cache (block#2867)
  fix(pairing): refuse unknown communities natively and name the requesting process (block#2835)
  Add per-interest Attention pause and management controls (block#2847)
  feat(drafts): show where unsent drafts wait (block#2853)
  feat(agents): let app-hosted agents read their owner's channels and members (block#2745)
  NEXUS-4271: Expose installed plugin status (block#2754)
  fix(menu): stop clipping lifecycle actions behind a hidden scrollbar (block#2716)
  feat(agents): give approved external apps the selected agent's full Buzz tools (block#2707)
  AIDA-2549: Keep Claude settings drafts in sync (block#2841)
  fix(channels): keep a channel's side panels open when returning to it (block#2838)
  fix(builderlab): clarify remote-agent account availability (block#2764)
  fix(sidebar): rename Remove from Team to Close conversation (block#2845)
  perf(channels): cut header and name-matching work in large channels (block#2726)
  fix(browser): silence a browser page once its tab goes away (block#2721)
  Community Mesh: shared compute plugin (including agent type) (block#941)
  perf(relay): cut main-thread work from profile updates (block#2730)
  fix(composer): lead the channel placeholder with its sidebar icon (block#2829)
  test(activity): scope preview assertions to the request row (block#2843)
  perf(search): fold people names once across palette keystrokes (block#2570)
  fix(github): keep PR watches running when localStorage is full (block#2699)

Signed-off-by: Cheng Peng <[email protected]>
Co-authored-by: Cursor <[email protected]>