[REQ-161][VERIFY] #857 稳态 catalog P95 复测（n≥10）+ 首启样本供 AC1b

[jinjunnn]

Parent / blocks

• Parent requirement: [REQ-161] 冷启动即用:模型目录及时可操作,无假「不可用」窗口 #528
• Closes latency face of: [REQ-161][BUG][P1] packaged 冷启动首轮 model.list 暴露 6,132 模型，P95 超过 2s #857 (correctness already landed via fix(ui-mac): bound the first governed model catalog (#857) #870 / fix(ui-mac): wake the first governed catalog read on catalog.updated (#882) #888 / [REQ-161][CODE] 首屏目录改为消费既有就绪事件，去掉 1.5s 自超时与指数退避 #882 / [REQ-161][DECIDE] 归因打包首启 sidecar-ready 之后的目录就绪成本 #881)
• Method source: docs/verification/2026-08-06-req109-110-t7-runtime/ (probe.ts)

Why

Owner 2026-08-10: AC1a = steady-state cold start P95 ≤ 5s, n ≥ 10; first-install samples excluded from that P95 and feed AC1b (value still TBD after #881 attribution). Prior T7 latency cells are stale (pre-#870/#888).

Outcome

On current main packaged macOS binary, produce a new evidence directory under docs/verification/ that either:

1. PASSes AC1a (n≥10 steady catalog-ready P95 ≤ 5000ms; first-list row count matches governed/hot path — no 6,132-row flash), and records first-install samples for AC1b; or
2. FAILs with measured numbers and a minimal follow-up CODE proposal (do not expand scope in this VERIFY).

In scope

• Build/sign packaged app from exact HEAD (same discipline as T7 README).
• Run probe in a quiescent desktop window only (open -g -j -n; stop if Dock flash would hit an owner-active session — return blocked with reason rather than contaminating the machine).
• Steady cells: logged-out/BYOK-only catalog-ready timings, n ≥ 10.
• Separate first-install samples (small n OK) for AC1b data only.
• Evidence README + results.json (+ screenshots if probe already emits them).
• Update [REQ-161][BUG][P1] packaged 冷启动首轮 model.list 暴露 6,132 模型，P95 超过 2s #857 comment with PASS/FAIL table; PR uses Fixes only if AC1a PASS and correctness still green.

Non-goals

• Rewriting product AC without owner.
• Re-opening [REQ-162][BUG][P1] token-only 换血等待 6s graceful stop，3s 续期档 catalog 约 10s #858/[REQ-162][BUG][P1] 快续期在 boot health 前 flush pending generation，单 fork 退化为 boot+换血 #859 token-rotation cells unless needed to get a clean BYOK steady sample.
• Deploying or mutating Project fields.

Acceptance

• Evidence dir cites exact git SHA + binary fingerprint.
• AC1a: n≥10 steady P95 ≤ 5s or explicit FAIL with P95 and sample list.
• First model.list row count equals stable governed set (not 6132).
• AC1b: first-install timings recorded; value left for owner.
• alpha-check / docs gate green for touched Markdown.

Boundary

alpha-code only. Local commits only until outer session publishes the draft PR.

[jinjunnn]

Claude session started (orchestrator)

• Worktree: alpha-code/.worktrees/verify-1053-catalog-p95
• Base: origin/alpha@229bd5e61669daafb8ad1d34734a0085c678eb92
• Session: aaabd924 (name verify-1053-catalog-p95) — local only; no push/PR until outer gate
• Constraint: quiescent desktop; return blocked if probe would interrupt owner

[jinjunnn]

Park — desktop not quiescent

Claude session ba7273d7 correctly blocked before build/sign/launch (worktree clean, no commit).

Evidence at check time (~2026-08-21 03:38 EDT):

• UserIsActive=1; HID idle ~5 min
• Frontmost Terminal; interactive apps open (Chrome, Cursor, WeChat, …)
• Production /Applications/alpha-code.app already running

T7 README forbids owner-active desktop probes. Resume when owner steps away / locks screen; then claude resume same worktree verify-1053-catalog-p95 @ 229bd5e61 (or rebase to current origin/alpha first).

Suggested reply from daemon: check again in ~20 minutes — orchestrator will not auto-spin; wait for explicit quiet window.

[jinjunnn]

Blocked — Owner action, review 2026-08-22.

quiet desktop for packaged catalog P95 probe

[jinjunnn]

Resumed — quiet-desktop handled by wait loop

Unparked / In progress. Claude session f53a9963 instructed to:

1. build+sign now (no Dock flash),
2. poll HID idle ≥15 min / screen lock (max 90 min) before probe launch,
3. then n≥10 evidence + local commit.

Owner authorized orchestrator to manage the quiet window this way (machine was still active at resume).

[jinjunnn]

Claude paused — session limit

Session f53a9963 blocked: Claude Pro session limit (resets 08:30 America/New_York). Progress saved in worktree.

Findings so far (not yet committed as VERIFY evidence)

• Quiet desktop / screen-lock path worked; packaged app built+signed (SHA-256 376a5c22… @ 229bd5e61).
• BYOK/logged-out cold-start probe reproduced twice: first model.list hangs through two ~10s client timeouts (~22s) before a later composer mount succeeds in ms once catalog_ready arrives → AC1a FAIL signal (gate ≤5s steady P95).
• Worktree dirty: docs/verification/2026-08-06-req109-110-t7-runtime/{probe.ts,results/results.json} (diagnostic + sample-count bump; full n≥10 evidence dir not landed yet).

Resume

After 08:30 ET: resume same worktree / session, finish diagnostic probe run, write evidence dir, local commit. Then outer session publishes draft PR (Refs/Fixes per PASS/FAIL).

[jinjunnn]

Blocked — Owner action, review 2026-08-21.

Claude session limit until 08:30 ET; BYOK ~22s hang already reproduced

[jinjunnn]

Resumed on Claude Max 20x

Owner upgraded quota. Session resumed (2be48062 / prior f53a9963 context) to finish diagnostic probe + evidence commit. Still In progress W34.

[jinjunnn]

VERIFY complete → draft PR + CODE follow-up

• Evidence commit 2c58b74eb — AC1a FAIL (P95 31,874.5 ms); row-count PASS
• Draft PR: [#1053][VERIFY] catalog-ready P95 recheck: FAIL 31.9s vs 5s gate #1055
• CODE follow-up: [REQ-161][CODE] 消除 home 双次 composer.mount：首链在 catalog_ready 前 10s 超时（#1053 FAIL） #1056 (dual composer.mount race)