fix(merge): preserve repeated H2 headings instead of dropping all but the last

[JeePeeTee]

Problem

blocksByKey builds a map keyed on the normalized heading, so a MEMORY.md that repeats an H2 — or has two that differ only in case, since keys are lowercased — collapses them into a single map entry. Merge then emits one block per key, so every section but the last is silently discarded.

This fires even when all three inputs are byte-identical. Merging a file that has two ## Notes sections against itself loses the first one's body. It reaches users through both callers: the claude-memmerge driver and Reconcile.

Solution

Parse disambiguates repeated headings by occurrence, so the Nth ## Notes in ours matches the Nth in theirs. The first occurrence keeps its plain key, so cross-document matching is unchanged and ## Deploy on one machine still matches ## deploy on another.

The separator is a NUL byte, which cannot occur in markdown, so a disambiguated key can never collide with a real heading's. Keys are internal to the merge algorithm and never appear in the output — only Block.Heading is written.

Verification

go build ./... && go vet ./... && gofmt -l . && go test ./... all clean.

Four new tests, each confirmed to fail when the fix is reverted:

• distinct keys at parse time
• both sections surviving an identical-input merge (got 1 in "# T\n\n## Notes\nsecond\n" before the fix)
• case-variant headings in one file both kept
• edits to different occurrences merging cleanly instead of one erasing the other

Found while reviewing the codebase; happy to adjust the approach if you'd prefer duplicates concatenated rather than kept separate.