0.36.4 2026-10-06

release 0.36.4 on GitHub →

Two improvements: the installer's drift gate stopped misreporting a correct scaffold as broken, and the native-boundary policy now documents why Zig is the named default instead of leaving it implicit.

Fixed

• gogogo --check reported a deliberately trimmed project as broken. The gate answers "would applyTrim succeed fully here?", so against a scaffolded tree every marker it looks for is legitimately absent — it emitted CHECK-FAIL for each removed unit and exited 1. A human knows to ignore that; an LLM agent reads exit 1 plus the word FAIL and concludes the scaffold is broken, which is exactly the wrong conclusion. Since apply already writes a Trim provenance section into the generated AGENTS.md, that information is now recorded as a machine-read removed: line and used to tell the two apart:

Verified end-to-end on a real scaffold: CHECK-OK 2 / CHECK-TRIMMED 5 / CHECK-FAIL 0, exit 0 (was exit 1 with 5 failures). Deleting a unit that was not recorded still exits 1 with the exact name — the gate still catches real drift, which is the property that must not regress. A tree with no removed: line is checked strictly, because a false "this is fine" is worse than a noisy failure. Unknown ids on the removed: line are ignored rather than trusted, since it is read from a markdown file a human may edit.

• The AGENTS.md template used backticks inside a Go raw string, which terminated the string literal. Caught by the compiler; the machine-read line is now plain text.

Changed

• docs/native-zig.md now documents why Zig is named rather than left open. The policy enforces "no native code without a Go baseline and a profile" — not "no preference about which toolchain". Those are different rules, and naming the default is what makes the 493-line policy, the zig-gate skill reference, and the boundary/removal plans pay off instead of being re-derived under pressure. The section cites the measured cost of choosing late from mini-redis-comparison: Mojo's discovery phase cost 120k prompt tokens versus 12.7k to reuse what it had learned (~9x), 4.96M total against Go's 387k; Odin's unprepared first submission did not compile (one hallucinated API); Zig on Linux was green 4/4, first try, 0 rounds, 0 hallucinated APIs. It also states plainly that Zig's stdlib churn is a cost, not a feature, and that a different language may win for a specific kernel — argued in required-justification item 5, never chosen silently.

Verification

• make ci-local green end-to-end; make site-check green (22 pages, all anchors resolve)
• Red-proofs: removing the provenance reclassification fails TestCheckAcceptsDeliberatelyTrimmedTree and TestExplainTrimCountsOnlyUnrecordedUnits; TestReadTrimProvenanceIgnoresUnknownIDs pins that an unknown id is not trusted; TestCheckStillFailsWithoutProvenance pins the strict fallback; TestWriteAgentsRecordsEmptyTrim pins that a full template still records provenance
• End-to-end against a real scaffold: trimmed tree exits 0, a unit deleted outside the recorded trim exits 1