Current status (2026-10-08 UTC)
Published as Chris directly to origin/main, without coauthors: 4e8ee8637, fix(ts): preserve ambient enum option and access diagnostics (#844).
The checker now reports ambient enum value import/re-export errors under verbatimModuleSyntax, and actual imported member/element accesses under isolatedModules. Unused isolated-only bindings remain legal; genuine TS2748 is preserved. Config verbatimModuleSyntax was parsed but dropped by the driver; it now reaches checker flags.
The original mismatch was independently verified on shallow source 30d3824: TS2748 at index.ts(2,31), while Home wrongly reported the import and re-export bindings. Deep missing-export failures were separate under #843. Initial failed oracle/config runs remain in evidence rather than being relabeled successful.
Verification and remaining acceptance
- All 24 no-flag / isolated / verbatim / both × unused / member / alias / element / namespace / type cases exactly match TS 6.0.3 diagnostic codes, counts, primary positions, messages and success/error exit classes.
- Six 1/8/32-layer valid/isolated cases also match, on combined published source
929d46awith #843's traversal. The 32-layer graph includes a cycle. - Full checker 4493, driver 199, Program 237, CLI 74 and entrypoint 33 tests passed; final Program/CLI/entrypoint reruns passed after the last #843 edge change. Harness 128 passed, no named test exclusions.
- Zod's complete 196 TS + 3 HM9002 diagnostic multiset is unchanged; this is not Zod admission. Pickier fails on the unchanged repository-wide 22,074 problems, not a lint pass.
- Keep open pending broader module-mode/ambient/import-equals and nested namespace acceptance review. Separate real
.tsdefault-import checking remains a failing oracle under #845; it is not among the 30 passing cases.
Documentation / costs
aed1dad3b publishes detailed docs, the brief README cost table and raw evidence: 592 pair rounds, 448 passing admissions, 6,206 verified checksums. Archive SHA-256 705bcf5818d80aaf4138ee9afc0efbe894af8049233dc2056a5f830ebf0e09cf.
Exact integrated-parent graph/scale confirmation medians are 2.9%/1.5% slower; both saving intervals span zero. Large-predicate 3.1% slower screen and 0.5% slower independent recheck are both retained. No speedup, equivalence or universal no-regression claim. Full report. Related #416, #843, #845, #548.