[feature] inspect: expose per-attempt model, timing, outcome and evidence #57

Split from #50 (item 6).

Problem

The store already records each attempt's harness, model_ref, started_at, ended_at and outcome (attempts table, crates/meshloop-adapters/src/store.rs:196), along with evidence rows. But inspect --json (crates/meshloop-cli/src/main.rs ~1210) only shows node state, liveness, worktree and note.

To see which model did the work, how long it took, how many retries it used, or what verification found, an operator has to query SQLite directly or read the harness transcripts. That was needed to estimate the cost of the #44/#45 run.

Proposal

• inspect --task <id> (and inspect --json per node, or behind an --attempts flag) lists each attempt with attempt_id, harness, model_ref, started_at, ended_at, duration_s and outcome, plus its evidence rows: kind, tool, exit, and summary_redacted or the path of the on-disk log.
• Add a run summary to status: total attempts, retries used, and total worker wall time.
• No schema change is expected. If one turns out to be needed, it follows the forward-only migration rule.

Relations

This is the read side for evidence that other issues add:

• [feature] persist worker stdout/stderr as attempt evidence so read-only nodes keep their findings #48 (worker output kept as evidence)
• [bug] empty verify_command silently passes nodes with no deterministic evidence #56 (verify: skipped rows)

Those issues can each write their rows independently. This one makes all rows visible in one place.

Acceptance criteria

• Fixture test: a node with two attempts → inspect --task lists both, each with harness, model_ref, timestamps, outcome and evidence rows.
• status --json includes an attempts and wall-time summary.
• No redacted field is shown unredacted.

🤖 Generated with Claude Code