mirror of
https://github.com/stablyai/orca.git
synced 2026-10-23 08:03:39 +00:00
Neil
06bf1bc4d3
docs/reference/ime-architecture.md says "Never hand-edit the bundles in the patch" and links to docs/reference/xterm-patch-regeneration.md. That doc does not exist, and neither does the harness it describes. Both landed in29117bf776and were deleted by17cfc968cf, a revert of the composition-ownership change, which swept up a build tool and a CI gate as collateral. The rule survived; its enforcement did not. Every xterm patch since has had to hand-edit minified bundles to comply with the surrounding architecture, because everything resolves to lib/xterm.mjs at runtime and under vitest, so a src-only edit is inert. The shipped bundles were therefore not the output of any build, and this restores them to build output. Comparing identifier multisets against a pristine build of the pinned commit finds hand-written names a minifier never emits ($rl, $hp, $tid), const in an otherwise let-only esbuild bundle, !! where the source reads Boolean(), an escaped LRM where esbuild emits the literal, and a return block esbuild collapses to void(...). Every remaining token difference is a minifier local reallocating. The old source patch could not be reused. It described the reverted composition-ownership architecture, so restoring it would have re-applied an abandoned design on top of dropping three accumulated fixes. It is re-derived from the shipped patch instead, and the derivation is a fixed point. Two deliberate departures from the deleted version. Sourcemaps are included rather than deleted, because a live test reads lib/*.map and asserts the mapped version matches the runtime version. The source-patch superset carve-out is gone, so a source hunk the shipped patch cannot name now fails loudly instead of being carved out silently. The doc's claim that the webgl and serialize addons reproduce byte for byte was half wrong. Their ESM output does reproduce at the pinned commit, but both also publish CJS that the root package script never builds, so folding either in needs a build step this harness lacks. Recorded as a blocker rather than a confident sentence. xterm_patch_sync runs the regenerator in --check mode, so a patch that does not match a rebuild of the pinned upstream now fails PR CI. The -diff -text attribute is required, not cosmetic: pnpm hashes the patch byte-for-byte, so a CRLF checkout breaks the install outright. Not verified: the CI job has not run on a real runner, the addon CJS bundles are unreproduced, and the generator is untested on Windows and Linux.