[roodboi]

Closes HACK-1319. This changes tests only: no product timeout or product code changes.

Findings

"moving job snapshot" fails locally (environmental, not a product defect). When the ticket was filed, this case's ordered start had a fixed 20 s deadline and a 30 s test limit. On the pre-#487 test at load average ~100, it timed out after 30 s; at ~58 it finished in 18 s of its 20 s deadline. Each fake Docker query is a cold Bun process, so the start was budgeting host process start-up, not product behavior. #495 (HACK-1452) already sizes that deadline from the counted freshness checks (ORDERED_START_FRESHNESS_CHECKS). On this branch's base, the case passed in every loaded full-file run below. This PR does not change it.

"real child" tests used wall-clock waits.

• The cancellation tests waited at most 15 s for the hanging fake Docker child to publish its PID. Under load the acquisition alone took ~14 s before its first probe, so the wait expired.
• The timeout test armed a 150 ms probe deadline. The probe reads Date.now() when it is created and spawns later, so a stall between the two could use up the whole deadline before any child existed (children.length > 0 failed).

Generation v5/v7 (and static7) cases outgrew their 20/30/60 s limits. With every limit lifted at load average ~90 (20 cores), the slowest cases took 27 s (plain saved adoption), 44 s (retained operations) and 60 s (public static7 up/down). The public adopted command gives each operation a 90 s deadline (DEFAULT_COMPOSE_STARTUP_TIMEOUT_MS), but those tests allowed 60 s in total. A Bun timeout leaves the owner running, so the file's settlement guard makes every later case refuse: one timeout produced 118 failures.

Changes

tests/native-compose-adoption-binding.test.ts

• startedChild(pending) is a handshake: it returns when the child publishes its PID, and fails with a distinct message if the operation settles first. It has no clock of its own. Both cancellation tests give their probe CANCELLED_CHILD_PROBE_DEADLINE_MS (60 s, the largest deadline the product admits), so only the abort ends it.
• The timeout test pins Date.now() (setSystemTime) while it acquires. The probe then arms its whole EXPIRING_PROBE_DEADLINE_MS (150 ms) as a real timer after the spawn, so there is always a child to reap. The clock is restored in finally, and the reaping assertions are unchanged.

tests/native-compose-adoption-generation.test.ts

• Named budgets replace the literals:
  • ADOPTION_CASE_TIMEOUT_MS: 120 s, for saved-adoption and CLI cases (~4× the measured worst).
  • RETAINED_OPERATION_DEADLINE_MS: the public command's 90 s deadline, replacing the 16 Date.now() + 20_000 caller deadlines.
  • retainedOperationCaseTimeout(n): adoption budget plus one deadline per operation, so a slow host ends in the product's own refusal before Bun's timeout.
• Expiry cases name EXPIRING_OPERATION_DEADLINE_MS (15 s, was 5 s, so the pre-callback probes fit) and DEADLINE_REFUSAL_SLACK_MS (5 s, was 1.5 s). The cancelled-effect bound is named CANCELLED_EFFECT_REFUSAL_BOUND_MS (10 s, was 4 s). It stays far below the 60 s fake child and the deadline, so only cancellation can satisfy it.
• Hanging fake children publish PIDs by rename (the HACK-1452 empty-file race, now fixed in this file too). The cancelled-effect start-up wait is a handshake instead of a 4 s poll.

Verification

• Linux host, Bun 1.4.2, synthetic CPU load (50 busy loops on 20 cores, plus ambient load): both files 5 consecutive times. In progress; this line will be updated with the results.
• Biome and tsc --noEmit: clean for both files.
• No macOS-specific behavior is involved. These fakes run on any host, so no M5 check is needed.

🤖 Generated with Claude Code

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

[linear-code]

HACK-1319

[roodboi]

Independent review (GPT-6-Astra, codex-dio): VERDICT: CLEAN at 75cf412. Reviewer thread mcp:cc4af3dc-6c8a-4e2e-9932-af0330123e26.

Static review of the exact-head diff against origin/next. GraphQL reviewThreads pagination returned no threads. The ordered signed merge from next and exact-head required test gate remain before READY. No PR merge performed.

[roodboi]

Independent review (GPT-6-Astra, codex-dio): VERDICT: CLEAN at 116509d. Reviewer thread mcp:cc4af3dc-6c8a-4e2e-9932-af0330123e26.

Re-reviewed after the verified signed merge of next at 472ec94. No manual conflict-resolution changes. Fresh paginated GraphQL reviewThreads returned zero threads. The independent review is static; hosted exact-head required test remains pending. No READY line until that check is green and the branch remains current. No PR merge performed.