Fewer things are called preflight

Performance

• policy-witness run no longer runs the unified-log deny scan on every run.
  That log show scan is archive-bound (~4 s on a large log store) and was the
  dominant per-run cost — not the sandbox evaluation. New --no-log-capture
  flag skips it: a small run drops ~5.6 s → ~0.08 s (~70×), and a full serial
  corpus pass goes from ~49 min to ~1.4 min.
• The host-side SBPL compile no longer runs on the happy path — only to
  disambiguate an xpc_error. Removes a redundant compile (~13 s on the 25 KB
  WebProcess profile).

Renamed — action required for callers

• sbpl-preflight → sbpl-check (it's a host-side SBPL compile-check, not a
  per-run preflight). Also: envelope kind sbpl_preflight → sbpl_check, and
  the run-envelope field data.policy_preflight → data.policy_check.

Behavior

• A non-compiling SBPL policy now surfaces as sandbox_apply_failed (from the
  worker), not the old controller bad_policy short-circuit. bad_policy in a
  run now comes only from the host's structural check (missing sbpl_source /
  non-sbpl format).

Tests & internals

• Foundational tests now witness the effect, not the prediction/shape
  (happy-path asserts the read succeeded; apply-isolation asserts
  sandboxed_after_apply; smoke asserts the observed deny).
• The menagerie now asserts the harness's emitted prediction + drift, and a
  new test exercises drift = true end-to-end by steering the validator's
  prediction — no live libsandbox bug required.

Migrating

• Binary: Contents/MacOS/sbpl-preflight → sbpl-check.
• Envelope readers: data.policy_preflight → data.policy_check; tool kind
sbpl_preflight → sbpl_check.
• Compile-error detection keyed on bad_policy should also accept
  sandbox_apply_failed.

---

Artifact: PolicyWitness-2.2.0.zip — signed, notarized, and stapled
(Developer ID; spctl accepted). Build CFBundleVersion 4.

EPERM enemy

Fixes a runner failure when sandbox_apply is disallowed (message-filter content and the like) and expands tests to include this case.

All-terrain vehicle

A new c worker allows us to manage version 2 and 3 profiles even with strict deny-default rules.

Attempt kinds are supplied for some common resources (e.g. file operations) and extendibility to other attempt kinds is offered.

Security, what a thing!

Make sure that you can only access the runner with debug attachments via the CLI, probably.

Too many things are called preflight

• Added a new host‑side SBPL preflight helper that compiles SBPL and emits a JSON envelope: controller/src/bin/sbpl-preflight.rs.
• Wired preflight into the controller run flow with fast‑fail on SBPL compile errors and extra diagnostics on XPC errors, plus new output fields: controller/src/run_flow.rs.
• Added preflight integration plumbing and documentation: controller/src/policy_preflight.rs, controller/src/main.rs, controller/Cargo.toml, controller/README.md, tests/build-evidence.py.
• Updated the build to build/copy/sign the new helper inside the app bundle: build.sh.
• Added signing guidance reminders for embedded helpers: AGENTS.md, SIGNING.md.
• Removed the unused runner selector helper and adjusted the test accordingly: controller/src/runner_select.rs.

Unification

Unified into a single runner with JSON output. End to end tests added to validated runtime witness. External "BYOSig" routine available for entitlement based restrictions.

new name, new me

• Renamed the distributable from EntitlementJail.app to PolicyWitness.app and the primary launcher from
  entitlement-jail to policy-witness (no backward-compatibility shims).
• Reset JSON envelope/witness schema_version values to 1 across Rust/Swift emitters, fixtures, and smoke
  assertions.