fix: harden end-to-end refactor workflow (#63-#70)

[caoergou]

This PR addresses the workflow gaps found during the end-to-end refactor run on xyz-data-website.

Changes

1. fix(engine): ignore owloop-owned paths in dirty-workspace check

   • .owloop/ is owloop's own state directory; it should not trigger the dirty-workspace prompt.
   • Closes Dirty-workspace check should ignore owloop-owned directories #63

2. fix(verification): honor → no output acceptance-criteria expectations

   • grep naturally returns exit code 1 when there are no matches. The verifier now treats → no output commands as passing when stdout is empty and the exit code is 0 or 1.
   • Closes Acceptance-criteria verifier should support semantic expectations like → no output #64

3. fix(report): auto-generate HTML report in worktree and fall back to session file

   • The engine now regenerates the static HTML report inside the worktree at the end of each run.
   • ReportGenerator falls back to session_latest.json when owloop_summary_latest.json is not present.
   • Closes HTML report should be regenerated automatically at the end of each run #65

4. fix(verification): restore excluded files mutated by acceptance-criteria commands

   • Commands such as uv run --with pytest can mutate tracked files listed in a spec's ## Exclusions (e.g. uv.lock). The verifier now reverts those files to HEAD after running acceptance criteria and backpressure.
   • Closes Prevent acceptance-criteria commands from mutating excluded files #66

5. feat(engine): add --no-push for review-before-push workflows

   • owloop run --no-push commits completed specs locally but skips the remote push, leaving commits on the branch for human review.
   • Closes Add --no-push / review-before-push option #67

6. feat(engine): include spec slug in owloop branch names

   • Generated worktree branches now embed the spec/task slug (e.g. owloop/20260707-extract-label-service-4a3b2c1) so it is clear what a branch is working on.
   • Closes Branch names should include the spec/task slug #68

7. feat(verification): classify functional vs meta-check failures as soft failures

   • When functional checks pass but a structural/meta check (e.g. grep ... → no output) fails, the run stops with a new soft_failure terminal state, preserves the diff for review, and avoids an unconditional rollback.
   • Closes Failure classification should distinguish code failures from meta-check failures #69

8. feat(engine): skip subagents for small/scoped specs to reduce token cost

   • --subagents now only enables the Orient/Implement/Verify subagent orchestration when a spec's ## Files scope exceeds the threshold (default 3 files). Small, scoped changes run as a single agent iteration to save coordination tokens.
   • Closes Reduce token cost for simple/scoped refactors #70

Verification

• uv run pytest -q: 381 passed
• uv run ruff check src/owloop tests: All checks passed
• uv run mypy src/owloop tests: Success