fix: CRLF-robust knowledge parsing + refuse malformed entry headers

[bilersan]

fix: CRLF-robust knowledge parsing + refuse malformed entry headers

Three read-side bugs, all observed live on Windows, where core.autocrlf=true
checks .context/ files out with CRLF line endings. The accepted entry-header
format is unchanged; no writer changes.

Spec: specs/fix-knowledge-crlf-and-malformed-headers.md (committed on-branch).

What ships

A. \r leaks into entry titles (internal/heading/entry.go, internal/heading/index.go)
Content is split on \n, so on CRLF files regex.EntryHeader's (.+) title
capture keeps the trailing \r. ctx disclosure inspect --json printed titles
like "Foo (consolidated)\r", and an apply plan written with clean titles
failed with "entry that is not in the root's staging zone". Fix: strings.TrimSpace
on the captured title at both read points; block lines are untouched.

B. False stale_header drift warnings (internal/drift/check.go)
checkTemplateHeaders compared the first comment of the live file and the
embedded template byte-for-byte, so a line-ending-only difference produced one
warning per knowledge file that ctx init --reset cannot fix. Fix: normalize
CRLF to LF inside extractFirstComment, which both sides go through.

build	CRLF files: stale_header warnings	LF files
upstream, LF-built binary	7	0
this branch, LF-built binary	0	0
upstream, binary built on a CRLF checkout (CRLF-embedded templates)	0	7
this branch, binary built on a CRLF checkout	0	0

C. Date-only entry header silently absorbed — data integrity (internal/disclosure)
ParseEntryBlocks only starts a block at a strict ## [YYYY-MM-DD-HHMMSS] Title
header, so a legacy/hand-written ## [2026-09-26] Title entry becomes part of
the previous entry's block: disclosure apply moved it into the previous
entry's theme file, and inspect never listed it. Fix: root validation (which
already refuses structurally malformed roots so the pass never mutates) now
refuses a root containing a ## [ heading that is not a full entry header, with
a typed MalformedEntryHeaderError{Line, Heading} (it also satisfies
errors.Is(err, ErrStagingUnparsable), so existing callers are unaffected).
HTML comments are skipped, so the ## [YYYY-MM-DD] Decision Title example inside
DECISIONS.md's <!-- DECISION FORMATS --> template does not trigger;
CONVENTIONS roots are exempt. Message:

progressive disclosure: line 32: "## [2026-09-26] Legacy" is not a full entry header; give it a full timestamp ("## [YYYY-MM-DD-HHMMSS] Title") so it is not folded into the entry above it

On the same file upstream reports "Moved 1 entries" and writes the Legacy entry
into the neighbour's theme file; this branch exits 1 with the root byte-identical.

Verification

• New tests: heading (CRLF titles for blocks and index), drift (the real
  LEARNINGS template in LF and CRLF, plus an edited header that must still warn),
  disclosure validate (date-only header in LF and CRLF with exact line/heading,
  the real DECISIONS template comment, a convention title containing ## [).
• golangci-lint v2.13.2 (go1.27.1): 0 issues.
• Full go test ./... on Windows: identical failure-name roster before and after
  every commit (pre-existing Windows-only failures: CRLF checkout, file
  permissions, raft cluster) — 0 new. The changed files add no LineLength /
  NoMagicValues findings.

Non-goal

ctx disclosure inspect is documented as never failing, so on a malformed root it
still lists the well-formed staging entries and leaves the malformed one out; only
apply refuses. Happy to follow up if you'd rather have inspect surface it too.

🤖 Generated with Claude Code