fix: read the YAML layouts a model actually writes, and name the line and construct on a parse error

[TheAmericanMaker]

Closes #246 (D-M10; mech 1.7, probes C/C2/D).

What the audit's probes actually show

Probes C and C2 (a sibling key indented deeper than the item's first key) are not valid YAML — PyYAML rejects both with "mapping values are not allowed here" — so the issue's "(valid YAML)" premise is wrong for those two. The layout the earlier itemIndent fix handles (- id: x with siblings aligned under id) already parsed. I probed a wider set of layouts against PyYAML and found four that are valid and were rejected or mangled:

Layout	Before	Now
List at the same column as its key (items: / - a)	items: null, and at the top level every later line silently dropped; nested, the indentation error	parses
Plain scalar wrapped onto more-indented lines (probe D — a long closeout_summary)	indentation error	folds like > (blank line → newline)
Scalar starting on the line after its key; wrapped sequence items; wrapped value on an item's first key	indentation / "mapping entry" errors	fold
Sequence of sequences (- - a)	indentation error	parses
Tab indentation	\tb: 1 → key "" (garbage)	error naming the line

Every expectation in the tests was checked against PyYAML before being pinned. Continuation stops at a mapping entry (key: — colon-space, so a wrapped https://… is not a key), a sequence item, or a comment; quoted scalars never continue.

Error messages

All parse errors are now YAML line N: <what was expected> — "<the line>":

• over-indented line: this line is indented 6 columns but the mapping it belongs to starts at column 4; a sibling key must align with the first key, and a wrapped value must not contain ": "
• dash where a key was expected (never valid inside a mapping's loop — the same-column-list case is consumed before the loop sees it): a sequence item where a mapping entry was expected; …
• content after the top-level block ended (was silently unread)
• tabs, duplicate keys (the Duplicate YAML key: dup wording is kept inside the new format — framework-handoff.test.mjs pins it), non-entry lines

The emitter is untouched; #225 (scalar-looking strings emitted bare) is a separate open high.

Docs

docs/library-format.md §YAML dialect still said > folded scalars are not read (stale since 0.19.3; L4). It now lists what is read and that errors name the line.

Tests

11 new (suite 732/732): 10 in tests/yaml.test.mjs covering each layout and each message, plus tests/handoff-layouts.test.mjs completing a real handoff written with a wrapped closeout_summary, a next-line description, and same-column lists through codecarto_complete. All 11 fail on the pre-fix parser.

🤖 Generated with Claude Code