perf(vapor): compile the three GB sdcc translation units in parallel

[lfkdsk]

What

buildGbRom compiled its three sdcc translation units serially (vapor/compiler/rom.ts). The compiler stage is the bulk of a GB build — the Pocket Vapor compiler itself is ~6% of wall clock, typescript.js load ~43%, external toolchains ~48% — so this runs the three units concurrently and keeps the link order unchanged.

Measured

	before (median)	after (median)	delta
full --target gb build	3108 ms	1847 ms	−40.6%
three-unit sdcc stage alone	2721 ms	1592 ms	−41.5%

Sample ranges do not overlap (serial min 3073 ms > parallel max 1882 ms). The stage is now bounded by the slowest unit (gen_app, ~1.6 s). On a single core the gain disappears (+1%), as expected.

Correctness

• All three .rel files, crt0.rel, the intermediate app.ihx and the final todo.gb (32768 B) are byte-identical to the serial build across 5 consecutive parallel builds.
• Parallelizing needed two failure-handling fixes beyond the scheduling change, because Promise.allSettled does not abort on the first failure the way the serial await loop did: failures are collected and re-thrown naming the unit, and the second commit (from the cross-family review) additionally removes partial .rel outputs and the stale final ROM on failure and preserves stdout-only diagnostics in link order.
• New vapor/tests/gb-build.test.ts with an sdcc shim (per-unit delay, partial-output and stdout-only failure injection); seven guard-removal mutations each fail a distinct case. bun test vapor/tests/: 77 pass / 0 fail.

Known limitation (pre-existing, not introduced here)

Two different apps building --target gb into the same output directory at the same time race on the shared gen-gb intermediates; this was observable before this change and is left for a follow-up.

Provenance

Autonomous fleet run: Scout-measured candidate (report on the fork branch lfkdsk/pocketjs:fleet/scout-compile-time), Builder implementation, cross-family review that hardened failure handling on top.

🤖 Generated with Claude Code

[lfkdsk]

Rebased onto current main (061c6be) and dropped the fork-only base commits, so the PR now contains only its own commit(s). bun test vapor/tests/ on the rebased branch matches main in this environment: everything passes except the NES parity build, which fails identically on unmodified main here (local ld65/cc65 library mismatch).