Tracework 0.13.4

Daily and Weekly guidance now keeps recorded decisions, delivery states, and future plans distinct, and shared storage writes avoid a Windows lock-initialization race.

Changes:

• Daily guidance explains supported behavior and choices before commits or test counts, preserves material rule conditions and exceptions, and treats missing measurements or reports as evidence limits rather than new required work.
• Weekly briefs use subsequent plans by default, reserve next-week commitments for supported dates, and keep undated actions, proposed acceptance criteria, and rejected suggestions in their recorded states.
• Shared JSON locking acquires the lock before initializing its byte region, preventing permission errors when concurrent Windows writers first create a lock file.

Compatibility:

• No configuration or vault migration is required; native installation remains tracework@tracework, and raw schemas and public commands are unchanged.

Validation:

• The complete CLI suite and generated-copy checks passed on Windows and GitHub Actions (Linux), both Claude manifests passed validation, and the tag dry-run passed from a clean isolated snapshot.
• Fresh isolated Codex and Claude installs matched all five skill directories byte for byte, and installed helpers resolved successfully; Claude validation retains its advisory that root CLAUDE.md is not automatically loaded as plugin context.
• The lock regression failed on the old code and passed on the fix; a real Windows contention check also confirmed that failed lock acquisition leaves existing JSON and lock bytes unchanged.
• In eight fixed-input writing cases, source review passed two original installed drafts, three original repository drafts, and seven final drafts under the same final review rules; one complex Daily draft still omitted material verification and delivery boundaries.
• Blind readability scores improved in four cases, declined in three, and tied in one, so a general readability gain is not established; these model and assistant reviews are not human acceptance or reruns of the reported engineering checks.