perf: environment scan tier, plus CI that can actually fail

[kaidesu]

What does this implement or fix?

Follow-up to #140. Two independent pieces: one performance fix for a workload shape the merged work regressed, and the CI gaps that let related breakage through unnoticed.

Added

• A middle storage tier in Environment, between the inline array and the map. Bindings are looked up in a fixed inline array, then a slice that only exists for environments that outgrow it, then a map past scanLimit.
• A TileRender benchmark: many globals in one flat scope read from a tight nested loop, with no function calls. The suite previously had no flat-scope case, which is why the regression below was missed.
• bench, fmt, vet and check targets in the Makefile, and matching CI steps.

Fixed

• A flat-scope regression from perf: amortize list.push and intern small integers #140. Inline environment storage made call-heavy code much faster but a tile rendering loop 15% slower. With four inline slots and a map behind them, a scope holding nine globals kept the first four inline and put the rest in the map, so reading any of those five paid a full scan and then a hash — strictly worse than the map it replaced. Raising the inline capacity fixes that scope and penalises every call frame, since the array sits in a struct allocated on every call; the two shapes of scope want opposite sizes, hence the middle tier. Tile rendering is ~8% faster than perf: amortize list.push and intern small integers #140, loop and map benchmarks improve, and the call benchmarks are unchanged.
• CI could not fail. make test pipes go test through sed for colour. A pipeline reports the exit status of its last command, so sed always succeeded and the target exited 0 no matter how the tests went — and the workflow runs make test. Verified by making a test fail and watching the exit code go from 0 to 1 once the target runs under bash with pipefail.
• The benchmark suite could not parse. ClassMethods still used Counter.new(), which Adopt JavaScript-style class syntax and fix OOP semantics #141 removed. Benchmark bodies are not compiled into a plain go test run and the Ghost programs inside them are only parsed when they execute, so tests stayed green while the suite was broken. Updated to new Counter(), and make bench now runs each benchmark for a single iteration in CI so this cannot recur silently.
• Workflow maintenance: actions/checkout and actions/setup-go move from v2 (retired Node runtime) to v4/v5; the Go version comes from go.mod rather than the ^1.17 that had drifted from the declared 1.21.1; pull_request is no longer filtered to the 1.0 base, so a PR is checked wherever it is targeted; the test timeout goes from 5s to 120s, since under -race on a cold runner the existing tests approach the old limit and risked flaking.

Removed

• The dep / Gopkg.toml bootstrap block in the workflow. There is no Gopkg.toml and modules have handled this since the repository moved to them.

Does this close any currently open issues?

None.

Additional Notes

On the performance number. The ~8% figure was measured against the pre-#141 tree. The class rework touched the evaluator enough that it is worth re-measuring on this base rather than carrying the number over; make bench is the starting point.

Why scanning wins. More than it looks like it should: the name being looked up and the name stored in the environment usually come from the same identifier in the AST, so the string comparison settles on equal pointers without examining any characters. Past scanLimit a scan genuinely stops paying, so everything migrates into the map and the scan tiers are abandoned — splitting names between a scan tier and a map is worse than either alone, which is exactly the bug being fixed. The tier is one slice of name/value pairs rather than two parallel slices; with two, the extra header made call-heavy benchmarks 10% slower even though they never allocate the tier.

Behaviour. Get, Set, Has, Delete, All, HasLocal and GetLocal keep their existing semantics; only the storage behind them changes.

Verification limits. gofmt, go vet, go build, go test -race and the benchmarks pass, but on every package except repl, whose liner dependency could not be fetched in the environment this was prepared in. CI covers the full ./..., and it is now able to report a failure if that is wrong.

Two separable commits. The CI fix is independent of the perf change and matters on its own, since the benchmark suite is broken on 1.0 today and CI is not reporting it. Happy to split it out if you would rather land that first.

Screenshots

N/A — no visual change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XtRdJPCskcS2ryiQZoEbUx

---

Generated by Claude Code