Updating the pinned test262 suite

The visible change is one line — SuiteGitSha in Jint.Tests.Test262/Test262Harness.settings.json. Treat it as a code change anyway. Upstream lands tests for normative changes that Jint has not implemented, and those turn green features red. The whole job is triaging the range before you run anything, then fixing the engine — never the tests.

Ground rules

• Never edit or "fix" a test262 test. If a test is wrong, the answer is an engine fix or, as a last resort, an ExcludedFiles entry.
• test262 at the pinned SHA outranks the published prose. See Normative changes.
• Exclusions outside intl402/ must be temporary and carry a comment naming what removes them. Today every ExcludedFiles entry is intl402 — keep that invariant.
• A feature that arrives unimplemented parks in ExcludedFeatures; the PR that implements it removes the entry.
• Work in a worktree, not the primary checkout, and start from a fresh main.

0. Orient

The runner does not read ../test262. State.Test262StreamLoader downloads the suite from GitHub at SuiteGitSha, so the local checkout exists only so you can read the range. Fetch it anyway — you cannot triage without it.

The GenerateTestSuite MSBuild target hashes the settings file into Generated/.settings.hash. Change any byte of the settings and the next build wipes Generated/ and regenerates. So: never pass --no-build, and if you suspect a stale generation, check the hash file rather than guessing.

1. Establish the range

cd ../test262 && git fetch upstream
git log --oneline <pinned-sha>..upstream/main            # or ..<target-sha> if one was given
git diff <pinned-sha>..upstream/main -- features.txt
git diff --name-only <pinned-sha>..upstream/main | grep -v '^test/'

Order commits by ancestry, never by author date. test262 merges rebased branches, so a commit authored earlier routinely lands on top of one authored later — a real bump had 3655e746 (authored 08-03) sitting above be13516f (authored 08-07). Use --pretty='%h %cd %s' or git merge-base --is-ancestor when the ordering matters, and confirm the true remote tip with git ls-remote https://github.com/tc39/test262.git main.

2. Triage every commit

Read the commits, not just the changed-file list.

For a small range, just read the new tests — git show <sha>:test/path/to/new-test.js. A test's info: block usually quotes the numbered algorithm verbatim, which is the fastest way to see what changed and the most reliable way to get argument-validation order right.

For each new behavioural test, check the engine before running the suite. Jint.Repl turns a 3-minute suite run into a 10-second answer:

dotnet run --project Jint.Repl -c Release -- -f scratch.js -t 10   # always pass -t

Write the assertions as plain if (…) throw new Error('FAIL …') — the REPL has no test262 harness. Strip the $DONOTEVALUATE() / includes: lines; for flags: [async] tests, assert on .then(...) callbacks.

3. Normative changes — the part that actually costs time

The published spec can be behind test262. A normative PR that reaches consensus gets test262 coverage while tc39.es/ecma262 still renders the old algorithm, and an already-merged proposal's own page (tc39.es/proposal-*/) is frozen at its pre-merge text and will confidently show you the wrong steps. Both happened on the Promise.try bump: the proposal page and the living spec agreed with each other and with Jint, and all three were what the new tests were written to reject.

Find the real change through the repo, not the rendered spec:

gh api -X GET search/issues -f q='repo:tc39/ecma262 <feature> <keyword>' \
  --jq '.items[] | "\(.number) \(.state) \(.title)"'
gh pr view <n> --repo tc39/ecma262 --json title,state,mergedAt,labels
gh pr diff <n> --repo tc39/ecma262        # the spec.html algorithm diff is the authority

Labels tell you the status: has consensus + needs implementations means implement it. mergedAt: null is not a reason to skip — AGENTS.md is explicit that where prose and test262 disagree, test262 at the pinned SHA wins.

When you implement one:

• Cite the document that owns the feature now. A merged proposal moves to https://tc39.es/ecma262/#sec-...; the anchors get renamed on the way in.
• Add a <remarks> naming the normative PR, so the next reader knows why the code and the current published prose disagree.
• Re-check the paths the reorder newly reaches. Moving a step can expose a latent bug that the old ordering made unreachable — see the gotchas below, where hoisting a call above a capability construction turned a dormant Promise.try() crash into a live one.

4. Bump, build, run

# edit SuiteGitSha, then:
dotnet test -c Release Jint.Tests.Test262/Jint.Tests.Test262.csproj

The build regenerates before it compiles; no separate generate step. ~3 minutes, ~99.9k tests. No runner timeout needed — the engine defaults to 30 s per test.

Verify the regeneration actually happened rather than trusting a green run over a stale Generated/:

grep -c "<new-test-name>" Jint.Tests.Test262/Generated/Tests262Harness.Tests.language.generated.cs

A negative parse test appears twice per file (strict + sloppy). Deleted upstream tests should be gone.

Then, because a spec fix moves engine behaviour that unit tests pin:

dotnet test -c Release Jint.Tests/Jint.Tests.csproj
dotnet test -c Release Jint.Tests.PublicInterface/Jint.Tests.PublicInterface.csproj

5. Fixing what fails

Failure output contains the failing script — strip the line numbers to reproduce it in Jint.Repl. Put a regression test in Jint.Tests for anything test262 does not cover (test262 tests the spec, not Jint's internals; a crash on an input no test262 test exercises still needs a pin).

6. Decide whether there is a PR at all

If the suite is green with only the pin moved, there is nothing to ship but a one-line pin bump — say so and ask before opening a PR. A bump that needs engine fixes is a real PR: pin + fix + regression tests, targeting sebastienros/jint main.

gh pr create --repo sebastienros/jint --head <you>:<branch> --base main

Keeping the local suite fork current is a separate, optional courtesy:

cd ../test262 && git fetch upstream && git merge --ff-only upstream/main   # then push origin main if wanted

Gotchas this task has actually hit

• Author date lies about ordering. %ad is not topology; use %cd or --is-ancestor.
• The proposal page is frozen at merge. tc39.es/proposal-promise-try/ still shows the pre-normative algorithm. Never cite it as current for a merged feature.
• spec.html is too big for a fetch tool. Asking a fetcher for one section of the living spec returns a truncated prefix and a confident "not found". Go through gh pr diff instead.
• A reorder exposes latent bugs. Promise.try moved NewPromiseCapability after the callback call; the argument-slicing that the old ordering never reached with zero arguments then crashed.
• arguments.AsSpan() is a trap in a built-in. ExpressionCache.ArgumentListEvaluation hands the callee Unsafe.As<JsValue[]>(object?[]) whenever the argument list is fully cached — including the zero-argument case — so AsSpan() throws ArrayTypeMismatchException ("Attempted to access an element as a type incompatible with the array"). Use arguments.Skip(n), which copies element-wise and returns [] when there is nothing to take.
• JsValue.Call on a non-callable throws a CLR ArgumentException, not a JS TypeError. Where the spec says Call(F, V, args), write GetCallable(f).Call(...) so the failure is a JavaScriptException your Completion handling can turn into a rejection.
• Promise-shaped fixes need the abrupt path checked too. Completion(...) wrapping means a throw inside the step becomes a rejection, while a throw from a step outside it propagates to the caller. Tests distinguish the two (ctx-non-ctor.js expects a synchronous TypeError; throws.js expects a rejection).