Engine — duplicate keys last-wins, inline-comment re-splice

Duplicates: Yams' duplicate-key error is disarmed by placeholder-
renaming first occurrences and re-composing (line count preserved, so
span layout stays exact); reads take the last occurrence, set()
rewrites the winner and clears earlier twins, remove() clears all.
Top-level only — nested duplicates stay YAML errors (doc updated).
Re-splice: an inline comment on a rewritten scalar line survives, with
quote-aware detection so a # inside quotes is never a comment. +31
tests (118 total).

Claude-Session: https://claude.ai/code/session_018BjQRYBR6jQja3jCRi5S3A

DESIGN/01-storage-format.md

@@ -81,7 +81,7 @@ Body: lane description / WIP policy / notes.
 
 Body: the card's content — the whole point.
 
-Schema-owned display fields are Lanework's to interpret — **coerce where a sensible reading exists, fall back to the field's default where none does** (settled). A scalar of the wrong YAML type reads as its source text (`title: 2048` displays as "2048", `width: "2"` reads as 2); where no sensible reading exists — a sequence or mapping where a scalar belongs, a non-integer width, an unparseable timestamp — the field falls back to its default: untitled placeholder, width 1, no color, no icon. Coercion is read-side only; the bytes on disk are **preserved verbatim, never rewritten**. One tombstone nuance: any *present* `deleted:` key tombstones the item — an unusable timestamp still deletes, its date merely unknown (the user's intent to delete outranks the broken date). **Duplicate keys: last one wins** (settled — the coercing read; strict YAML would reject the file, so this is a deliberate divergence in the editor's favor): a key appearing twice reads as its last occurrence, earlier occurrences preserved verbatim on disk and invisible. An app write of a duplicated key rewrites the winning (last) occurrence and removes the earlier ones — the app owns the keys it writes, and leaving a stale twin would resurrect it if the winner were later removed; removing a key removes all its occurrences. **App rewrites preserve comments** (settled): comments on their own lines always survive a rewrite; an inline comment on a rewritten value line is re-spliced after the new value — best-effort, guaranteed for plain scalar lines (the realistic case), dropped only in pathological shapes. Fail-fast remains reserved for structure (`schema`, `order`, YAML validity) — and it covers *malformed*, not just missing: an `order` that is present but non-numeric is the same loud malformed-input rejection as a missing one. **Readable-but-uneditable shapes load; writes to them refuse** (settled): frontmatter the surgical editor can't key by spans — a whole-frontmatter flow mapping, non-scalar keys — reads and renders normally, and any app write to that file fails loudly through the per-file write-failure banner (02-architecture.md ▸ Write-failure surfacing) naming the shape, never a silent corruption and never a load rejection.
+Schema-owned display fields are Lanework's to interpret — **coerce where a sensible reading exists, fall back to the field's default where none does** (settled). A scalar of the wrong YAML type reads as its source text (`title: 2048` displays as "2048", `width: "2"` reads as 2); where no sensible reading exists — a sequence or mapping where a scalar belongs, a non-integer width, an unparseable timestamp — the field falls back to its default: untitled placeholder, width 1, no color, no icon. Coercion is read-side only; the bytes on disk are **preserved verbatim, never rewritten**. One tombstone nuance: any *present* `deleted:` key tombstones the item — an unusable timestamp still deletes, its date merely unknown (the user's intent to delete outranks the broken date). **Duplicate keys: last one wins** (settled — the coercing read; strict YAML would reject the file, so this is a deliberate divergence in the editor's favor): a *top-level* key appearing twice reads as its last occurrence, earlier occurrences preserved verbatim on disk and invisible. A duplicate inside a nested mapping value remains a YAML error — the rescue applies where the editor's slip actually happens, the top level. An app write of a duplicated key rewrites the winning (last) occurrence and removes the earlier ones — the app owns the keys it writes, and leaving a stale twin would resurrect it if the winner were later removed; removing a key removes all its occurrences. **App rewrites preserve comments** (settled): comments on their own lines always survive a rewrite; an inline comment on a rewritten value line is re-spliced after the new value — best-effort, guaranteed for plain scalar lines (the realistic case), dropped only in pathological shapes. Fail-fast remains reserved for structure (`schema`, `order`, YAML validity) — and it covers *malformed*, not just missing: an `order` that is present but non-numeric is the same loud malformed-input rejection as a missing one. **Readable-but-uneditable shapes load; writes to them refuse** (settled): frontmatter the surgical editor can't key by spans — a whole-frontmatter flow mapping, non-scalar keys — reads and renders normally, and any app write to that file fails loudly through the per-file write-failure banner (02-architecture.md ▸ Write-failure surfacing) naming the shape, never a silent corruption and never a load rejection.
 
 ## Enhanced schema (reserved, out of scope)