perf(native): project-local content-keyed transform cache, lazy @babel/core

[danfry1]

Problem

Three cost sinks in the native engine's transform pipeline, all from the audit:

1. The transform disk cache lived in os.tmpdir() — ephemeral on CI runners, so every job paid the full cold Babel transform of React Native's ~250-file boot graph, and macOS purges it periodically, silently re-cooling local caches.
2. Disk entries were keyed by path + mtime + size. Fresh installs and Docker mtime normalization invalidated everything (or worse: a same-size, same-mtime file with different content — RN's paths are version-independent — could serve stale executable code).
3. @babel/core was loaded eagerly before the cache lookup. Requiring Babel costs ~35ms vs ~0.5ms for resolve-only version reads; under the default engine (isolate: true) every worker paid that even with a fully warm disk cache.

Change

• Cache moves to <projectRoot>/node_modules/.cache/vitest-native/ (tmpdir fallback when node_modules is absent or unwritable — strictly more environments work than before). The V8 compile cache is colocated so both share a lifecycle, and the location is restorable by standard CI cache actions (wired up in ci: cancel superseded runs, cache bun downloads and transform output, bound matrix cells #72; the two PRs land in either order).
• Disk key = sha1(platform + path + content); the directory name carries the preset version, @babel/core version, and the Babel environment. The path and env are in the key deliberately: Babel's output is not content-pure — the preset's dev-mode JSX transform embeds the absolute filename (_jsxFileName) and differs under NODE_ENV=development vs test/production (verified empirically against @react-native/babel-preset 0.85). A pure content key would share file A's embedded identity with file B. CI-restore validity is what matters and is preserved: CI workspaces use a stable checkout path and env. The content hash replaces a statSync (the source string is already in hand at both call sites), so the hot path gains no I/O.
• @babel/core loads lazily on the first cache miss; init() only resolves paths and reads versions, and the friendly install-guidance error still fires at startup for the missing-package case.

Measured

Package's own native suite (29 files), warm cache: aggregate worker setup −30% (4.4s → 3.0s), wall clock −11% (1.75s → 1.56s) — scales with test-file count since every isolated worker pays init. On CI the relocation eliminates the per-job cold transform entirely once #72's cache step is active.

Review

Independently adversarially reviewed pre-PR (no blockers). Findings incorporated: exact-key test assertions replacing racy directory-entry counts; the filename/env non-purity was caught during development and is regression-locked by the key scheme and tests. Known accepted nits: no cache eviction yet (entries are small; follow-up), and a corrupt-but-resolvable Babel install now errors at first miss rather than startup (the raw error is arguably more actionable than wrong install advice).

Full gate: mock 1297 + 9 new cache tests, native suites (stock/hot/isolation/android/navigation/soak), crosscheck 75/75, consumers from the packed tarball, lint/format/typecheck clean.