test: make generation preparation responsiveness deterministic

[IchenDEV]

Outcome

Release v0.0.54 failed because its preparation responsiveness test could finish
copying before the heartbeat task ran. Replace the timing probe with the existing
worker entry/exit barrier. The MainActor must respond while real preparation is
held at that barrier; copied bytes and candidate cleanup are still checked.
Internal callbacks default to nil for existing production callers.

SDLC bundle and risk

• Bundle: docs/sdlc/changes/2026-10-16-preparation-responsiveness-test/
• Risk: medium, test reliability and an internal instrumentation seam.
• Owner explicitly approved continuous design, execution, verification and merge
  after checks pass. A future production run requires separate approval.
• Failed v0.0.53 and v0.0.54 tags remain unchanged. The v0.0.54 first failure and
  skipped production are preserved in the upstream signing bundle.

Verification

• bash scripts/sdlc-checks.sh and bash scripts/ci-basic-checks.sh.
• Protected swift test --skip-build --skip-update --disable-automatic-resolution
  after building outside the runtime policy with native SwiftPM sandboxing.
  1,143 unique XCTest cases, 17 conditional skips, zero failures; one Swift Testing
  case passed. Every started/finished identity matches frozen discovery.
• Complete native-engine execution of the same 14 frozen bundles also passed.
  All 702 product files and 300 bundle files retain their hashes.
• Forcing the entire worker onto MainActor caused exactly the intended
  assertion failure without hanging. Exact source restored, rebuilt and passed.
• Release-style Developer ID app build, resource checks, DMG checksum and
  mounted-content verification passed. This local fixture is not notarized or published.
• Independent local code/evidence/signature review passed.
• Runtime privacy canaries, unchanged 19 dependency pins, and public-content
  credential/path scan passed. Raw logs remain private.

No UI change. Initial wrapper environment failures and the manifest prewarm
resolution are retained in the committed evidence, with native sandboxing and
privacy controls preserved.

Residual risk and rollback

PR CI and exact-main CI must pass before a new release tag. Production signing,
Apple notarization, Gatekeeper checks and publication remain separate release
outcomes. Revert this scoped commit through a PR to roll back; preserve the
responsiveness requirement and failed immutable tags.

Reviewer focus

The entry callback, actual preparation and deferred exit callback must remain in
the same detached closure. The MainActor snapshot must precede gate release.