fix: status honours --config, observed liveness for unowned workers, record skipped verification

[smota]

Summary

Implements #46, #47 and #56. Closes #46, closes #47, closes #56.

• [bug] status ignores --config and reports ok with empty runs while a run is active #46: status now parses and honours --config and --db. A missing config, or a --graph that doesn't resolve, returns ok: false and exits 2. Before, both returned ok: true with no runs.
• [bug] inspect reports live: Dead for detached-run workers that are running #47: status/inspect used to show live: Dead for workers started by a different Meshloop process. They now fall back to the attempt's saved PID and image name, the same check crash_fail already uses. Recovery deliberately keeps the old check, where Dead means "this process can't collect the worker", so orphaned workers are still killed and retried (feat: contain harness descendants on posix when meshloop aborts #30).
  • Live testing turned up a second bug, also fixed here: the attempt's saved image_name was the harness name ("claude") rather than the process image (pwsh.exe). So no saved attempt could ever match its process. Harnesses now report process_image() from their configured executable.
• [bug] empty verify_command silently passes nodes with no deterministic evidence #56: an empty verify_command used to add no evidence, so a node advanced as if it had been checked. It now records a verify: skipped evidence row, doctor --json reports it under warnings, and run prints a warning to stderr.

How this was built

The work was delivered through Meshloop: graph issues-46-47-56 (nodes 1 and 2), then issues-56. The orchestrator was Claude Code (Opus 5.5) and the workers were Claude Sonnet. You accepted both plans.

Commit	Issue	Worker	Orchestrator fixes
78579d8	#47	node 2, attempt 4	cargo fmt
07be0ee	#46	node 1, attempt 3	Removed a duplicated #[test]. An unknown --graph actually fails in the saga.status error path, which still printed only to stderr, so that path now returns a JSON error too.
eab9828	#56	issues-56 node 1, attempt 5	Moved a const that had split #[allow(clippy::too_many_arguments)] from its function
8808951	#47	— (orchestrator)	Store the worker's process image name instead of the harness name

All three worker attempts failed Meshloop verification. In each case the logic was right, but there were format or lint problems that cargo fmt or clippy would have caught. The worker log for attempt 5 shows why: the worker command allowed "Bash(cargo:*)", but Claude Code still denied cargo ("requires approval"), so no worker could run fmt, clippy or its tests. I salvaged each attempt's diff onto this branch, made the fixes listed above, and re-ran the workspace check. These attempts never reached Meshloop's technical acceptance. This PR is the review point.

Checks

• cargo run --locked -p xtask -- check (fmt, check, clippy -D warnings, workspace tests, xtask scaffold) passes on 8808951. Windows 11.
• New tests:
  • parser: status keeps --config/--db/--graph
  • CLI: status with the config outside the current directory; no config; unknown graph
  • engine: six observed_live cases; skipped_verify_evidence
  • CLI: doctor warnings with empty and non-empty verify_command
  • adapters: process_image_name; a spawned child with the right image reads Live, and with the harness name reads Ambiguous
• Live, against the detached issues-56 run: status --config run from another directory listed the running node. With no config, it returned ok: false.
• Not run: a live inspect of a running detached worker after 8808951. While that worker was running, the binary built before the image-name fix showed Ambiguous, which is the correct result for a mismatched image. The new adapter test covers the matching-image case.
• Not run: Linux/macOS. PosixProcessView doesn't compare image names, so a reused PID on POSIX can still read Live. That was already the case for crash_fail.

Limitations

• attempts.ended_at is never written, so observed_live's finished-attempt check doesn't fire in practice. Finished workers fall through to the PID check, which reports Dead or Ambiguous correctly.
• Plain meshloop with no arguments runs status. With no config it now exits 2 with "no config found" instead of printing the banner.

Follow-ups, filed separately, are linked in the comments.

🤖 Generated with Claude Code

[smota]

Follow-ups found during this run: #64 (verify evidence drops the error), #65 (redaction false positive in worktree paths), #66 (ended_at never written), #67 (second ready node not dispatched while one is verifying), and a comment on #27 (worker tool rules didn't let workers run cargo).