fix(#2703): strip GSD-2 frontmatter with the canonical parser (#3027)

* test(#2703): failing-first coverage for CRLF frontmatter strip in SUMMARY.md

Drives the exported buildPlanningArtifacts seam. Rows for CRLF/LF parity,
stacked blocks and a leading BOM fail against the current hand-rolled
regex; the negative-space rows pin behavior that must not change.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(#2703): strip GSD-2 frontmatter with the canonical parser

buildSummaryMd matched the closing delimiter with a hardcoded bare \n, so a
CRLF-authored task summary never matched and fell through to the raw-passthrough
branch. The function then prepended its own block, emitting a SUMMARY.md with two
stacked frontmatter blocks and no warning.

Delegates to stripFrontmatter from frontmatter.cts -- the canonical, line-ending
tolerant primitive this repo already deduplicated once (#2143) -- instead of
adding another hand-rolled variant.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(#2703): strip only the first frontmatter block in gsd2 import

Adversarial review caught a regression in the first cut: stripFrontmatter
loops by design, so a summary body opening with a thematic-break-delimited
section (--- / heading / ---) had that section silently deleted. The old
pre-#2703 regex preserved it, so shipping the loop would have traded one
silent corruption for another.

Adds an explicit { once } option to the canonical primitive -- default
behavior and the two existing callers are unchanged -- and has buildSummaryMd
opt in. A GSD-2 summary is an arbitrary user document, not a GSD artifact with
a known doubling failure mode, so a second block there is body content.

This also makes the acceptance criterion exact: CRLF now produces the same
result LF already produced, rather than a new result for both.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* chore(#2703): backfill changeset pr number

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

src/frontmatter.cts

@@ -642,22 +642,29 @@ const FRONTMATTER_SCHEMAS: Record<string, { required: string[]; requiredValues?:
 };
 
 /**
- * Strip ALL frontmatter blocks from the start of `content`.
+ * Strip frontmatter blocks from the start of `content`.
  *
- * Handles CRLF line endings and multiple stacked blocks (corruption
- * recovery): greedily strips consecutive `---...---` blocks separated by
- * optional whitespace, so a doubled/tripled frontmatter header (e.g. from a
- * botched merge) is fully removed, not just the first block.
+ * Handles CRLF line endings and, by default, multiple stacked blocks
+ * (corruption recovery): greedily strips consecutive `---...---` blocks
+ * separated by optional whitespace, so a doubled/tripled frontmatter header
+ * (e.g. from a botched merge) is fully removed, not just the first block.
+ *
+ * Pass `{ once: true }` to stop after the first block. Callers whose input is
+ * an arbitrary user-authored document — rather than a GSD artefact with a
+ * known doubling failure mode — need this: a body that opens with a
+ * thematic-break-delimited section is lexically indistinguishable from a
+ * second frontmatter block, and the greedy loop deletes it silently (#2703).
  *
  * Canonical home for this primitive (#2143 audit dedup): previously
  * duplicated byte-identically in both `state.cts` and `state-transition.cts`.
  */
-function stripFrontmatter(content: string): string {
+function stripFrontmatter(content: string, opts: { once?: boolean } = {}): string {
   let result = content;
   while (true) {
     const stripped = result.replace(/^\s*---\r?\n[\s\S]*?\r?\n---\s*/, '');
     if (stripped === result) break;
     result = stripped;
+    if (opts.once) break;
   }
   return result;
 }