fix(conformance): async-safe docker probes + real imageDigests staleness

[pofallon]

Conformance runner: async-safe Docker probes + real imageDigests staleness

Two follow-up fixes from the /code-review of the 022 declarative conformance runner (findings #4 and #5). Both are contained to the dev-only parity-harness crate.

#4 — no blocking std::process::Command on the async executor (Principle V)

The Docker channel observers each spawned their own blocking docker inspect from a sync capture() invoked inside async fns, blocking the tokio executor.

• The runner now pre-fetches the container's full docker inspect once, off the executor via tokio::task::spawn_blocking, and stores it on RunContext.container_inspect.
• The four Docker observers (image/process-graph/injected-process/temporal) became pure readers of that pre-fetched value — no subprocess, no blocking.
• The workspace-container lookup in execute_ops is likewise offloaded via spawn_blocking.

Net effect: a Docker case now pays one inspect instead of 4+, and no observer blocks the async executor. spawn_blocking works on the current-thread runtime #[tokio::test] uses. The one-shot refresh / snapshot check CLIs and the Drop cleanup guard keep direct sync calls (non-concurrent, the sanctioned exception).

#5 — imageDigests is now a real staleness signal

imageDigests was declared a staleness field but was never populated or recomputed — an image-content change on a future Docker snapshot case would go undetected.

• New image_digests_for_case: for a Docker case, recomputes the digest of each image its fixtures pin (docker image inspect → RepoDigests/.Id); for a non-Docker case it returns empty without touching docker (a read-configuration snapshot must never gate on a base-image digest — the same false-staleness trap avoided for host tool versions in BUG: Features installed in running container instead of during image build #1).
• Wired into both the record path (capture_provenance) and the live replay gate (snapshot_oracle, offloaded). The hermetic snapshot check (no docker) keeps carrying the recorded digests — it cannot verify a Docker case's images, so it must not falsely flag them (returns None → carry recorded).

The only existing snapshot case (case-readconfig-snapshot) is config-only, so its imageDigests stays empty and the committed snapshot is unchanged.

Verification

• cargo fmt --all -- --check + cargo clippy --all-targets --all-features -- -D warnings clean.
• Full hermetic conformance + parity suite green, incl. the real-Docker channel tests (now exercising the pre-fetch path) and new unit tests for image_digests_for_case/fixture_image.
• Live parity_conformance_runner passes against real Docker + the pinned oracle.

🤖 Generated with Claude Code

https://claude.ai/code/session_019vhr7Tcf8ybvSZSE1owqBT