feat: merge undersized heading sections during chunking (chunking.min_chunk_size)

[j-sperling]

Summary

• Heading-dense markdown (conversation transcripts, session logs — memsearch's primary memory shape) currently produces one chunk per heading section no matter how tiny; a 19,195-session corpus chunked into 186,192 pieces (~10x expansion), and embedding cost scales linearly with chunk count. This adds chunking.min_chunk_size: consecutive sections below the threshold are accumulated into one chunk (never exceeding max_chunk_size), carrying the first section's heading and the full merged line span.
• Default 0 = exactly today's behavior (every section flushes immediately); the merge path only activates when opted in. Oversized sections still route through the existing paragraph splitter unchanged.
• Plumbing: ChunkingConfig.min_chunk_size (TOML + config set), allowed in project-local .memsearch.toml (same trust tier as the existing chunking.max_chunk_size/overlap_lines), passed through MemSearch.__init__ and the CLI override map.
• Trade-off note for review: merging changes chunk identity (chunk IDs hash content + line span), so enabling it on an existing collection re-embeds affected files once. Retrieval-quality impact is corpus-dependent; on a LongMemEval-S bench harness, coarser session-level granularity scored slightly better than per-turn chunks with the same embedder, but a maintainer-side eval would be worth running before recommending a non-zero default.

Test plan

• New tests: default keeps per-section behavior; tiny sections merge (heading/line-span/order assertions); merge respects max_chunk_size; pending chunks flush before an oversized section; accumulation stops once the threshold is reached
• Updated _cfg_to_memsearch_kwargs translation test and manual-construction test stubs for the new attribute
• Full suite: 241 passed, 7 skipped
• ruff check / ruff format --check clean

Design notes

• Merging is heading-level-blind by design. Accumulation crosses heading levels (a small ### tail can merge into the following ## section). For the transcript shape that motivates this feature, levels carry no hierarchy (every turn is the same level), and stopping at level boundaries would reintroduce the tiny-chunk problem exactly where it bites. If level-aware merging turns out to matter for prose docs, it can be layered on later without changing the config surface.
• Merged chunks keep the first section's heading in metadata only. All subsequent headings survive verbatim inside content (sections include their heading lines), so no text is lost to search; the metadata heading is used for display/filtering and the first section's heading is the least surprising representative for a run of consecutive tiny sections.