chunking: edge guard hardcodes 5 bases instead of using kmer_context

[jayhesselberth]

Found during the generalization audit (#125). Not tRNA-specific, but the same class of baked-in default.

Problem

The chunk edge guard hardcodes 5 bases rather than using the configured kmer_context:

• src/leech/chunking/extractor.py:350 — base_idx = int(np.clip(base_idx, 5, leech_read.num_bases - 6))
• src/leech/chunking/extractor.py:360 — focus_bases = list(range(5, leech_read.num_bases - 5))
• src/leech/preparation/parallel.py:142-144 — return list(range(5, max(5, num_bases - 5)))

5 is DEFAULT_KMER_CONTEXT, but ChunkConfig.kmer_context is configurable. A run with kmer_context=10 still guards only 5 bases, so edge chunks get N-padded k-mers; a run with kmer_context=2 needlessly discards usable focus bases.

Same pattern, lower stakes

• src/leech/chunking/serialization.py:80-81 — chunk.get("feature_start", -5)
• src/leech/preparation/parallel.py:275-276 — config.chunk.feature_start or -5

The second one has an additional bug: or treats a legitimate feature_start=0 as falsy and silently substitutes -5. Should be an explicit is None check.

Suggested fix

Use chunk_config.kmer_context at all three guard sites (np.clip(base_idx, k, num_bases - k - 1), range(k, num_bases - k)), thread it into _find_motif_positions, and derive the feature_start/feature_end fallbacks from it as well.