fix(frontmatter): warn on comment loss, and lift rationale into the body

[evemcgivern]

Frontmatter comments cannot survive a write. This makes the loss loud, and gives rationale a home that survives.

The bug

lib/frontmatter.py round-trips frontmatter through JSON:

meta = _yaml_to_dict(match.group(1))   # yq -o=json .
yaml_text = _dict_to_yaml(meta)        # yq -P .  <- json.dumps(meta)

JSON has no comment concept, so every writer erases every frontmatter comment — refresh-md, reconcile, slot, hygiene, anything reaching write_file. Structurally, not occasionally.

A routine hygiene --yes run deleted 213 lines of ranking rationale from a real track. The next_up order survived perfectly — all 45 entries, correct sequence — which is exactly what made it invisible: a sanity check on the ordering passes, and the loss only appears in a full diff. The data survived; the reasoning did not.

Two changes

1. write_file warns, naming the count. count_frontmatter_comments is fail-soft: an unreadable or frontmatter-less file has nothing to lose, and a warning path must never be able to break a write.

2. lift-rationale moves rationale into a ## Ranking rationale body section. The body is passed through write_file verbatim, so it survives — and there's a test pinning that premise rather than assuming it.

Why separate instead of preserving comments

Preserving them via targeted yq -i writes was the obvious alternative, and it was rejected for a reason beyond effort: frontmatter comments never render as markdown, never reach the VS Code viewer, and never appear in export --json. Rationale kept there was already invisible to every surface anyone actually reads. The fragility just made that visible.

The part most worth reviewing

Association is by indentation, not adjacency.

# TIER 1 — the golden path is broken     <- indent 2, header
- 6374   # first line, inline            <- entry, indent 2
         # continues here                <- indent 9, belongs to 6374
- 6368   # ...

Adjacency cannot distinguish a trailing comment from a section header — both sit between two entries. My first implementation used adjacency and glued TIER 2 banners onto the preceding issue; test_header_is_not_glued_to_the_previous_entry is that bug, pinned.

The inline comment on the entry line also has to be captured: the real format puts the first and most important line of rationale there, so a naive - (\d+)\s*(?:#.*)?$ silently drops the summary of every entry. test_inline_comment_on_the_entry_line_is_captured covers it.

Tests are written against the real file's shape, not an invented one — TestRealFormat is transcribed from the track this was found on.

Evidence

Claim	Result
python3 -m unittest discover tests	1474 tests, OK
tests.test_lift_rationale	25 tests, OK
Dry run, --repo=CritForge	170 comment lines / 72 blocks on one track, 11 / 5 on another — correctly grouped
Clean track produces no warning	test_write_is_silent_when_there_is_nothing_to_lose
Body survives a write verbatim	test_body_survives_a_write_verbatim (incl. # and : in the prose)

The falsification cases matter as much as the positive ones here: a warning that fires on every routine write becomes noise and gets ignored, so a comment-free track must stay silent.

Notes

• Pure stdlib, offline tests, per CLAUDE.md.
• README updated: new reference row plus a warning on the weekly-cadence line.
• Does not change how frontmatter is serialized — that stays a JSON round-trip. This makes the consequence visible and gives rationale somewhere durable, rather than pretending YAML comments are safe.

Refs #491