fix: make local semantic-line-breaks consume CI's NLB checker (closes #2085)

[dem-extra1]

Closes #2085.

The local reformatter (scripts/semantic-line-breaks.py) and CI's check-new-line-breaks job disagreed on mid-line semicolons (and on lowercase-follower sentence boundaries). --write could therefore produce a CI failure. Measured on #2073 at e88f2eec and again on #2092.

Construction

Agreement is by construction, not by two implementations happening to match:

• .github/workflows/validate.yml already pins Morrison-Lab/gha/check-new-line-breaks@fdaeccc47273fb8c542d05b00d488242b8563b15.
• scripts/vendor/gha-check-new-line-breaks.py is that script at that SHA (byte-identical; do not hand-edit).
• scripts/vendor/gha-check-new-line-breaks.pin records the git SHA and the sha256 of those bytes.
• scripts/lib/nlb_gate.py refuses to load if the pin disagrees with validate.yml or the file bytes.
• The reformatter emits through emit_gate_clean: sentence splits come from the gate's split_sentences; clause splits run only when classify_line returns clause.

Refresh after an action-pin bump: python3 scripts/sync-nlb-checker.py.

Predicates checked (not only the semicolon)

Against the vendored checker at that SHA, the local splitter also missed _SENT_BREAK_LOWER_RE (a period after two lowercase letters, then a lowercase word). That path is now the gate's function, so --write splits it and no longer rejoins a hand-break. Extra lowercase abbrevs (min/hr/hrs) ride along on that same branch.

#2081 left alone

The gate does not flag comma clauses (gha#336). A long comma-clause sentence stays one line; scripts/test_slb.py pins that. Closing the 200-400 character comma-join trap is a different issue.

Tests

python3 scripts/test_slb.py: 79 passed. Covers the pin, the #2085 semicolon line, lowercase-follower splits, code-span masking on a line long enough to be a clause, ellipsis-before-capital, and that #2081 comma joins are unchanged.

  

[dem-extra1]

Session wrap 2026-08-26 12:31 PDT: the Ready on 271a6f4 still stands and mergeable_state was clean. The non-blocking _fetch FileNotFoundError finding was not Addressed this session (Cloud Agents blocked on Cursor monthly quota). Tracked as #2338. Do not merge from this comment.

Posted by Claude Code (AI agent) --- not written by a human.

[d-morrison]

Pushed 233eba9 addressing the adversarial finding raised against 271a6f4 (local/CI agreement was coincidental --- the gate pinned the checker's SHA but ignored the action's clause-config inputs):

Addressed. nlb_gate.py now parses validate.yml's NLB step with: block (clause-breaks/clause-min-length) via a new resolve_nlb_config(), resolving both through the vendored checker's own _env_flag/_env_int for exact CI parity, and emit_gate_clean defaults to and threads that resolution through every recursive classify_line call instead of the compiled-in defaults. Today's live config still resolves to (True, 80), and a synthetic-validate.yml test proves the reviewer's example line ("Short clause here; second bit.") flips from clean to a clause violation, and from left-whole to split, once clause-min-length drops to 10. The module docstring's "by construction" claim now states that config, not just script identity, is derived from validate.yml. 86/86 tests pass (79 prior + 7 new).

The one earlier finding (sync-nlb-checker.py FileNotFoundError without gh) remains Deferred to #2338, as before.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[d-morrison]

Pushed 980922d addressing the round at 233eba9:

Addressed --- the unguarded module-scope import yaml is gone. The import now lives inside parse_ci_nlb_with, guarded with the repo's friendly-exit convention (same pattern as validate-skills.py), so every yaml-free path --- --help included --- works on a machine without PyYAML, and a run that genuinely needs CI-config resolution exits with the pip install pyyaml instruction instead of a raw traceback. Two regression tests inject a blocking yaml shim on PYTHONPATH and assert both directions. 88/88 tests pass.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[d-morrison]

Pushed 112dd70 addressing the adversarial round at 233eba9:

#	Finding	Disposition
1	Config cache never engaged on the real path (~170x measured whole-file regression --- validate.yml re-parsed per sentence)	Addressed --- emit_gate_clean now hands load_nlb_config the caller's checker (None on the ordinary path), so the no-override cache holds; a counting test pins resolution to at most once across repeated calls; whole-file preview of CLAUDE.md measured back to 0.14s.
2	clause-breaks axis had no non-default regression coverage (dropping it passed all tests)	Addressed --- synthetic clause-breaks: 'false' config test added: it resolves, flips classification of a long semicolon line, and contrasts against the live default.
3	Unguarded module-scope yaml import	Addressed earlier in 980922d (lazy guarded import, friendly install message, shim-blocked regression tests) --- confirmed by that round's note that the head had advanced.

Now 92/92 tests. The interim CI failure at 980922d was the new PyYAML test depending on a git base ref CI cannot resolve; fixed in cf65603 (probe via --all on a temp file).

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[d-morrison]

Pushed 750e14c addressing the final adversarial delta round at 112dd70:

#	Finding	Disposition
1	The guard's sys.exit escaped the reformatter's per-path except Exception loop (SystemExit is not an Exception), so a multi-path run without PyYAML died at the first path with no per-path errors and no summary	Addressed --- the guard now raises NLBPinError, which the per-path loop reports before continuing to the summary; a multi-path regression test asserts each path reports and Done ( still prints.
2	The comment claimed convention parity with validate-skills.py's module-scope guard, whose clean-exit guarantee a lazy per-sentence guard cannot give	Addressed --- comment rewritten to state the actual contract and why the exception form is used here.
3	Dead REPO_ROOT binding in test_slb.py	Addressed --- removed.

93/93 tests. That round also verified no cache pollution from synthetic-config tests, explicit-config operation with yaml fully absent, and the perf fix (0.136s vs 9.921s pre-fix, byte-identical output).

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.

_{2026-08-26 23:01 UTC}

[github-actions]

Claude finished review — View run

---

Correcting this: I ran scripts/check-pr-fully-clean.py, which is the correct enumeration instrument (not gh pr checks), and it returned exit 1 with real finding bullets — this PR is not currently fully clean:

❌ PR is NOT fully clean:
  - Check run 'review / claude-review' is still in status 'in_progress'
  - Check run 'build / build' is still in status 'in_progress'
  - Check run 'validate' is still in status 'in_progress'
  - No review comment has been posted evaluating HEAD SHA 750e14c5 yet

My prior message's headline ("Ready for merge, pending CI completion") was self-contradictory — a PR with in-progress required checks and no verdict-bearing review comment at its current HEAD is not ready for merge, full stop, regardless of qualifying language attached afterward. Restating accurately:

Code-level findings: unchanged from my analysis above — the new commit (750e14c5) correctly fixes the SystemExit-escaping-the-per-path-loop regression, is tested (93/93 local), and introduces no new defects. That part of the review stands.

PR status: NOT fully clean. build, validate, and review/claude-review are still in_progress on HEAD 750e14c5, and no review comment has yet evaluated this exact HEAD SHA (this review itself, once posted, will be the first). Do not merge until those checks complete green and a verdict-bearing review lands at this HEAD.

Verdict

Blocked / not clean at time of writing — CI is still in progress on HEAD 750e14c5 and no review comment yet evaluates this HEAD, per check-pr-fully-clean.py (exit 1). No new code-level defects found in the delta since the prior round (750e14c5 correctly fixes the batch-killing SystemExit regression, tested and verified). Re-check CI/check-run state once it settles before merging.

Reviewed commit: 750e14c

Stopping Point: Clean stopping point reached — this review-only task is complete. My prior message's "Ready for merge" framing was inaccurate and is superseded by this correction; no further review-only action is queued.

Reviewed commit: 750e14c

[github-actions]

💰 Cost: $3.2861 (review) — run