The engine-split soak, dead-end, live-hold, landing-simulation and OAuth browser test files run under the test budget #12391

I fixed four of the five slow test files and added a nightly run for the long soak. landsim has not changed: I couldn't find a way to speed it up. The summary checkbox was already done: scripts/test-shards.ts:978 prints the five slowest tests and their share of the run, using slowestTestLines in scripts/test-budget.ts. Times below were measured in this sandbox, which is heavily loaded (load average about 90).

• test/live-hold.test.ts: 31.6 s → 1.6 s, and it leaves SLOW. One test waited out the client's 30 s gap between page pulls in real time. The client now runs on a fake clock: a timer moves the clock on by its delay and fires at once. The held → swapped test still exercises that gap.
• test/engine-split-soak.test.ts: 64 s → about 3–5 s a test, so it stays in SLOW with a new reason. The gate now runs 48 round trips (three per client) at once, beside a writer that never pauses, so a stale read or a gap in the stream still fails it. The 1000-trip soak is a new step in the nightly outside-gate job, so test/nightly.test.ts now lists that step too (src/nightly/jobs.ts, docs/features/nightly.md). It stays over 2 s mainly because the first requests stall about 4 s while competing with the writer.
• test/deadends.test.ts: 75 s → 39 s, measured back to back. It stays in SLOW with a new reason.

- Its seven walks (three seeds' crawls, the near-miss pass, and three offered-action passes) now each run on their own forge in a worker thread (test/support/deadends-worker.ts), side by side. - A crawl now fetches eight pages at a time but takes them from its queue in order and judges them in order. It therefore visits what the one-at-a-time walk did, and the KNOWN list is unchanged. - The worker imports src/origins/mirror.ts first, because loading the modules in another order hits a "used before initialization" error.

• test/oauth-connect-browser.test.ts: likely fixed, but not verified. Its "Continue" link opens a new tab, so clickAndFollow waited the full 15 s deadline for a page load that never came. It now stops as soon as the consent screen is reached. The test also now skips, as test/signin-asking.test.ts does, where the Chromium it finds cannot start. Chromium can't start in this sandbox, so I kept its SLOW entry. If the check reports it as "in SLOW, yet nothing over the budget", the entry can be removed. There is no separate lane for Chromium tests, and I didn't create one.
• test/landsim.test.ts: unchanged. Two attempts failed: running each seed in a worker made it slower (24 s against 11 s), and caching git lookups by object id made no difference. Most of its time is in seeds 1, 2 and 4, which run the order, merge and determinism checks. Seed 2 alone takes about 4.7 s.

scripts/test-times.json lists none of these files, so I left it alone. I couldn't update the plan on the issue: the keep_plan tool isn't available to me. I made no time promises, so nothing was recorded with estimate.

bun scripts/gate.ts and bun x tsc --noEmit pass. scripts/fast.ts named 1720 files, too many to run, so I ran the 57 files that touch SLOW, the nightly jobs or test/helpers/chrome.ts, plus the changed test files: 346 pass, 92 skip (the Chromium tests), 0 fail.

🤖 Generated with Claude Code

Four of the five slow test files are now faster, and the long runs moved to the nightly outside-gate job. Only test/live-hold.test.ts leaves SLOW. The other four stay in SLOW with rewritten reasons, because a single test in each still takes over 2 s. The summary checkbox was already done: scripts/test-shards.ts:978 prints the five slowest tests and their share of the run (slowestTestLines). The sandbox was heavily loaded (load average 80–350), so local times are rough.

• test/live-hold.test.ts: 31.6 s → 1.6 s, leaves SLOW. One test waited out the client's 30 s gap between page pulls in real time. The client now runs on a fake clock: a timer moves the clock on by its delay and fires at once. The held → swapped test still covers that gap.
• test/engine-split-soak.test.ts: 64 s → 3–5 s a test. The gate runs 48 round trips beside a writer that never pauses, so a stale read or a gap in the stream still fails it. The nightly job runs the 1000-trip soak.
• test/landsim.test.ts: 32 s → 12 s here. The gate runs 4 seeds at once instead of 6. Seeds 1, 2 and 4 still run the additive-merge, determinism and order-independence checks. A new test fails if the quick seeds stop covering any of the three. The nightly job runs 24 seeds through the test file, which fails on any violation no issue owns; the command-line simulator always exits 0, so it isn't used for this.
• test/deadends.test.ts: 75 s → about 36 s.

- The near-miss and offered-action walks now run only with DEADENDS_ACTIONS=1 (or DEADENDS_LONG), which the nightly job sets. Every KNOWN finding comes from the crawls or the sign-in checks, which stay in the gate, so that list is still checked in full. - The gate keeps the three seeds' crawls. Each crawl takes 18–22 s, alone or with the other two in parallel, so dropping a seed would not shorten the file. - Seed 2 is the only one that shows two of the KNOWN findings. - The crawls now run in worker threads and fetch eight pages at a time, but judge them in queue order, so they visit the same pages as before.

• test/oauth-connect-browser.test.ts: still over 2 s a test, kept in SLOW. The 15.8 s was clickAndFollow waiting out its 15 s deadline after "Continue" opened a new tab. It now stops once the consent screen is reached.

- The repository has no separate lane for Chromium tests. This test now skips when Chromium can't start, as the other browser tests do. - The whole-gate check on 736e39f did not print "in SLOW, yet nothing over the budget" for this file, so it is still over 2 s. That log doesn't give its actual time.

What to look at:

• src/nightly/jobs.ts: the three new outside-gate steps. test/nightly.test.ts and docs/features/nightly.md list them too.
• scripts/gate.ts: the rewritten SLOW reasons.
• test/support/deadends-worker.ts: the deadends worker.

bun scripts/gate.ts and bun x tsc --noEmit pass. I ran test/landsim.test.ts, test/nightly.test.ts and test/deadends.test.ts: 31 pass, 2 skip (the action walks), 0 fail. The earlier run covered the 57 files that touch SLOW, the nightly jobs or test/helpers/chrome.ts. Chromium can't start in this sandbox, so the OAuth test only skipped here. To try the nightly runs: LANDSIM_SEEDS=24 bun test ./test/landsim.test.ts and DEADENDS_ACTIONS=1 bun test ./test/deadends.test.ts.

🤖 Generated with Claude Code

Closes #12382

These changes take the slow suites from #12382 out of the gate's path or speed them up. What the gate covers stays the same; the long runs go to the nightly job.

• Engine split soak: the gate does 48 round trips. The nightly job outside the gate does the full 1000.
• live-hold: the live client runs on a fake clock, so the 30 s gap between pages passes without waiting. The test is no longer listed in SLOW.
• clickAndFollow: stops waiting once it reaches what it follows. A step that opens a new tab no longer sits out the 15 s deadline waiting for a load event.
• deadends: walks its seven forges side by side, one per worker thread, and a crawl asks for eight pages at once. The near-miss walks and offered-action walks now run nightly (DEADENDS_ACTIONS=1) instead of in the gate.
• landsim: the gate runs its four quick seeds at once; the 24-seed corpus runs nightly.
• SLOW: each test still listed there now says why it is slow.

The last commit rewraps two comment lines the gate's 100-column rule rejected, in src/nightly/jobs.ts and test/landsim.test.ts. It changes nothing else.

To review: the job entries in src/nightly/jobs.ts, which name what moved to the nightly run, and the worker split in deadends.

Here, bun scripts/gate.ts and bun x tsc --noEmit pass, and so do test/landsim.test.ts and test/nightly.test.ts. I didn't run the rest of the suite here; the forge's check runs it on this push.

🤖 Generated with Claude Code

Since main can't go into soak tests, this branch splits up the slow tests that kept the gate waiting. The engine-split soak now runs 48 write-then-read round trips beside a second engine process writing as fast as it can. The full 1000 run in the nightly job outside the gate, and the engine starts in a worker thread. test/deadends.test.ts crawls its seeded forges side by side, each in its own worker thread, and asks for eight pages at once. Its near-miss and offered-action walks run nightly. In the gate, landsim lands four seeded scenarios at once; the nightly job runs 24. clickAndFollow stops waiting as soon as it reaches what it follows, so a step that opens a tab no longer sits out the 15 s deadline for a load event. That stall was what kept test/live-hold.test.ts on the slow list, and it is off it now. Each test still on the slow list in scripts/gate.ts says why it is slow.

This push merges main in. The only conflict was the landsim line in SLOW in scripts/gate.ts: main (#12119) pointed it at issue 12447, and this branch rewrote its reason. The merged line keeps this branch's reason and main's issue number.

No behaviour changed apart from that line. bun scripts/gate.ts and bun x tsc --noEmit are clean. scripts/fast.ts could not name the tests because the changes merged in from main make the diff too broad. So I ran the gate-list tests (check-tiers, check-list, check-allow-lists) and landsim myself: 84 passed, 0 failed.

To try it: bun scripts/gate.ts, then bun test test/landsim.test.ts test/engine-split-soak.test.ts test/deadends.test.ts.

🤖 Generated with Claude Code

This branch splits the slow tests that held the gate. Each heavy test runs a short pass in the gate and its full run in the nightly outside-gate job (src/nightly/jobs.ts, docs/features/nightly.md):

• Engine-split soak (test/engine-split-soak.test.ts): 48 write-then-read round trips beside a second engine process writing as fast as it can. The nightly job runs the full 1000.
• Dead ends (test/deadends.test.ts, test/support/deadends-*.ts): the gate crawls the seeded forges side by side, each in its own worker thread, and asks for eight pages at once. The near-miss and offered-action walks run nightly.
• Landing simulator (test/landsim.test.ts): the gate lands four seeds at once; the nightly job runs 24.
• Live hold (test/live-hold.test.ts): it runs on a fake clock and is off the slow list. clickAndFollow in test/helpers/chrome.ts now stops waiting as soon as it reaches what it follows, so a step that opens a tab no longer waits out the 15 s deadline for a load event.

Each test still on SLOW in scripts/gate.ts says why it is slow. The merge with main kept this branch's landsim reason and main's issue number (12447).

The red check: "an engine that stops saying it is alive is ended and started again" was flaky, and it fails on main too when run alone in this sandbox. On a loaded box an engine Worker takes about 2.5 s to boot and sends nothing until it is up. The test's 1.5 s stall window ended each engine before it wrote its first row. The test decides "first life, so hang" from those rows, so every restarted engine hung the same way, and engine.ready() never resolved. The fix is in the test only: stallMs goes from 1.5 s to 5 s, still well under the 15 s default, and the wait for the first restart from 15 s to 20 s. test/engine-split.test.ts passed 12 of 12 three times in a row, with exactly one restart.

Not done here: src/engine/host.ts has two weaknesses, both left alone because fixing them changes how the engine behaves:

• ready() returns the promise of the current launch, so a caller already waiting when the engine restarts waits forever.
• The stall watch counts from launch, so a boot longer than STALL_MS (for example a long migration on a big database) would loop restarts in production too.

bun scripts/gate.ts and bun x tsc --noEmit are clean. scripts/fast.ts couldn't name the tests (the diff is too broad), so I ran engine-split, landsim, check-tiers, check-list and check-allow-lists myself; all pass.

To try it: bun test test/engine-split.test.ts test/engine-split-soak.test.ts test/landsim.test.ts test/deadends.test.ts.

🤖 Generated with Claude Code