test(startup): 45s load budget for full-server-boot launcher tests (#542)

[lis186]

摘要

修 #542：三條 launcher-mode 測試在「某台機器一致失敗、另一台一致通過」——判別特徵是它們正好是在 spawnAndCollect 8 秒上限內完整啟動 server 並斷言 exit 0 的三條測試。與 #538 同類（83204be 在重現機上實測飽和負載下 server boot ~31s，無載 ~1s），持續性背景負載（如背景 review 工作）讓它們連 isolation 跑都必然超時。五個同類站點預算 8s→45s，對齊 83204be。

Diagnosis

• The discriminating feature: of all spawnAndCollect call sites, the three failing tests (plus two structural siblings) are the ones that boot a full server inside the cap and assert on its completed launch. Arg-validation and loop-guard sites exit before boot and are unaffected.
• Three launcher-mode tests fail on one machine, pass on another #542's "fails in isolation too" observation is consistent with load, not evidence against it: the load on the reproducing machine was sustained (background review jobs), so isolation doesn't relieve it. 83204be (the test: index-fields.e2e importer case fails locally on main (expected 1 index line, found 0) #538 fix, same machine, same week) measured boot at ~31,097ms under 14-way saturation — 4× the 8s budget.
• Local quantification on this 12-core machine (which does not reproduce): under 12-way busy-loop saturation the three tests consume 5.3s / 0.4s / 6.4s of the 8s budget — 80% erosion in the mildest form of the same conditions; 24-way oversubscription stays under the cap because macOS scheduling protects the foreground work.

The change

Five sites in test/startup.test.js move 8000 → 45000 (83204be's budget):

• the three Three launcher-mode tests fail on one machine, pass on another #542 tests (spawns claude through the provider registry, consumes --no-browser, launches codex app on macOS)
• the hub-discovery sibling (same spawn shape, lighter path)
• the E2: claude not found test (boots the same server before reaching the ENOENT)

All five child processes exit on their own on the happy path (ENOENT path calls finish(1) — verified in server/index.js:644-652), so the timeout is a backstop, not a wait: a passing run's duration is unchanged.

Verification

• Fixed branch, test/startup.test.js complete file: 64/64 pass.
• Full suite CCXRAY_HOME=$(mktemp -d) npm test: 2190/2190 pass, exit 0.
• Honest limitation, per docs/verification-principles.md fallback: a local fail-on-old is not achievable on this machine (the failure needs the reproducing machine's sustained load profile; busy loops here max out at 80% budget erosion). The differential evidence for this failure class is 83204be's instrumented 1/7 fail → 7/7 pass under 14-way saturation on the machine that reproduced both test: index-fields.e2e importer case fails locally on main (expected 1 index line, found 0) #538 and Three launcher-mode tests fail on one machine, pass on another #542.

Review gate

Trivial test-budget change (5 numbers, in-repo precedent 83204be, no production code touched) — submitted without a grok pass; say the word if you want one anyway.

Fixes #542

🤖 Generated with Claude Code