Mechanical cleanup: single-source checksum, exit_status, run layout, timestamps, version

[qinhaihong-red]

Parent

#2

What to build

Verified small redundancies from the walking-skeleton review, batched as one cleanup pass. Each item should be resolved or explicitly declined in the PR description:

• definition checksum is computed twice per run (and the workflow is model-dumped three times); build the snapshot once and reuse its checksum
• exit_status is persisted twice in the same attempt record (column + inside the normalized output JSON); keep a single durable source
• Run-directory layout literals are split across the CLI (runs root) and executor (file names) with tests re-hardcoding both; give the layout one owner
• UTC-ISO timestamp construction is duplicated between the events module and the executor; share one helper
• The event log reopens its file on every append; hold one append handle for the log's lifetime
• Workflow.version is required, unconstrained (any int validates), and never read; constrain it to known versions or drop it
• The package is fully typed but ships no py.typed marker, so downstream type-checkers degrade caw types to Any
• version duplicates the project version from pyproject and is never read; single-source it (e.g. importlib.metadata) or remove it

Acceptance criteria

• Snapshot/checksum computed once per run
• exit_status has a single durable source
• Run-directory layout owned in one place
• Shared UTC timestamp helper
• Event log holds one append handle
• Workflow.version constrained or removed
• py.typed shipped
• version single-sourced or removed

Blocked by

None - can start immediately

[qinhaihong-red]

Partially addressed by #90 (the PR for #12). The "single-source run layout" item now has a home: src/caw/runlayout.py exposes runs_root() / run_dir(run_id), and caw run / caw resume / caw report all resolve .caw/runs/<id> through it — the three inline Path.cwd() / ".caw" / "runs" literals in cli.py are gone.

Remaining for this issue: the in-run layout is still spelled inline in executor.py — the state.sqlite / events.jsonl / workflow.normalized.json filenames and the runs_root / run_id join at the write sites. Folding those into runlayout.py (e.g. state_path / events_path / snapshot_path, and having execute_run derive its run_dir from the helper) finishes the single-source so there is exactly one authority for the layout.