[zibi-pplx]

Summary

About 70 of the pplx-space-api tests listed in ci/test-duration-budget.toml get faster. Each one still proves what it proved before, and production keeps today's values. The slow tests fell into four kinds.

The test Kubernetes substrate always waited out its budgets. Its reflectors are seeded once and never fed, so in tests:

• every pod-cache miss waited the full 250 ms reflector budget;
• every delete waited the 1 s blocking-delete timeout;
• every follower of a joined restore saw its owner's change only on the next 2 s restore poll.

The reflector wait and the retry-deferral grace move into SubstrateBudgets, and the restore poll ceiling becomes a field of the restore waiters. The test substrate uses 20 ms, 20 ms and 100 ms. with_seeded_caches, the public test constructor that cluster-controller tests use, keeps production values.

Production bounds the tests waited out. These become injectable; production passes the existing constants:

Bound	Production	Tests
Sandbox list page byte budget (SandboxController.list_page_max_bytes)	8 MiB	KiB-sized pages; the same 1/7/8 page shapes
Quota lock timeout (PublicOperationAdmission.quota_lock_timeout)	2 s	400 ms
Builder connect retry (DirectSubstrate.builder_connect_retry_policy)	the default TCP policy	200 ms

Real waits in the tests themselves.

• The native pod failover test, which proves a caller waits out the full 10 s connect budget, runs on tokio's paused clock. Its 9 s assertion still holds, in virtual time.
• Fixed 1 s sleeps that showed a request blocked on a lock now poll pg_stat_activity for the waiter. The poll helper moved from the quota tests to test_support.rs.
• Lock and access timeouts that a test waits out are shorter. Each still sits above the backoffs and sub-cases it has to separate.
• The update-timeout test's lock observer kept running until its deadline after the update had finished. It now stops with the update.

Sequential sweeps. These loops become rstest cases:

• state rendering, as 8 snapshot shards. The union of their rows equals the old snapshot, and each shard still generates the full lifecycle list.
• hidden operations, e2b connect, sessions, owner scope, admin list and id forms;
• the Kubernetes store behaviours, now 100 tests, each with its own store;
• lifecycle characterization, native create policy, and create-ending-early.

The file-write tests also build each multi-MiB body once instead of re-encoding it.

Test	Before	After (slowest case)
routes::e2b::state_render_tests::each_surface_renders_…	15.8 s	0.38 s (8 shards)
native_build::a_failed_system_build_frees_its_key_…	10.6 s	0.38 s
hidden_operations_answer_as_nonexistent_…	10.6 s	0.41 s
updates::postgres_tests::unchanged_timeout_does_not_wait_…	9.8 s	0.32 s
characterization_tests::kubernetes::completed_actions_keep_…	6.8 s	0.67 s
the_write_body_limit_admits_the_largest_documented_payload	5.7 s	0.49 s
owner_scope_tests::scoped_sandbox_page_bounds_the_bytes_it_returns	5.4 s	0.37 s
restoring_access::a_joined_restore_that_its_owner_abandons_…	4.6 s	0.69 s

These are local times on a loaded machine.

Stack

1. #1407, #1408, #1414, #1417, #1422: the test duration budget and the changes so far.
2. This PR: pplx-space-api's slow tests.
3. The same work for the other crates follows on top of this PR.

Validation

• cargo clippy -p pplx-space-api --all-targets --all-features -- -D warnings and cargo fmt are clean.
• Locally, 2,498 of the crate's 2,502 tests pass. The other four are native_pod_failover tests: they bind 127.0.0.2, which macOS does not allow, and fail the same way without this change.
• The paused-clock failover test passed in 0.18 s locally on a routable stand-in address; CI on Linux runs it as written.
• This PR does not yet remove exceptions. Once CI shows which listed tests now run well under the limit, I will remove theirs. Tests that became rstest cases have new names, so their old entries go too.

Severity and risk

Severity: does not apply; this is a CI speedup.

Risk:

• Production. The only changes are values moving from constants into fields and parameters. Every production construction path sets them to the same constants.
• Flakiness. Shorter test waits leave less slack under load. Each shortened wait still exceeds the backoffs and sub-cases its test separates, and the split ineligible-row cases got a longer lock-observation deadline (500 ms) because they now start at once.
• To roll back, revert this PR.

AI assistance: Claude Code drafted the change, its tests and this description, using parallel agents whose patches it combined.

🤖 Generated with Claude Code

[coderabbitai]

Caution

Review failed

The pull request is closed.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

[zibi-pplx]

Combined into #1426 with the rest of the test burn-down, so the stack is #1407, #1408 and #1426.