[emerose]

Summary

• Latent bugfix in SPEC-CHUNK-322. The "encourage split before a heading" discount keyed on _is_heading (a standalone heading chunklet). Stage 2 almost always glues a heading to its first sentences, so that predicate is rarely true on real prose — the discount was effectively dead and chunks packed across section boundaries.
• The split-before discount now keys on a new _starts_with_heading predicate (first block token is a heading). The split-after forbid stays on _is_heading, so a lone heading is still never stranded at a chunk's tail.
• Bumps to 0.6.2 and updates the spec + changelog.

Before / after (Qasper coherence set)

Paper	Before	After	Effect
1908.05925	12 chunks, 9 start at a heading	14 chunks, 13 start at a heading	removes a stranded glued-heading tail; the 1646c equation chunk is preserved byte-identical
1909.13375	17 chunks	19 chunks	fixes a mid-sentence tail and a chunk that packed past ## Model ::: Objective and Training; formula stays intact
1910.07601	16 chunks	16 (unchanged)	n/a

Every new cut lands on a section boundary; smallest new chunk is 784c (above the ~410c general-term floor), so no tiny stubs.

What this deliberately does NOT do

• No cross-heading penalty (over-split in testing — flat ## rendering can't distinguish major sections from siblings).
• No under-target penalty — small chunks are fine when the document is genuinely structured that way; short-but-coherent sections are left intact.

Test plan

• uv run pytest — 214 passed, 20 skipped
• ruff + pyright clean
• New test_split_before_glued_heading_with_body proven to fail on the old keying and pass with the change
• Equation-keeping, stranded-heading prevention, tiny-stub avoidance all still hold

🤖 Generated with Claude Code