Port Astra macsyfinder filename fixes to Aksha

[jwestrob]

Context

Astra's --write_macsyfinder flag produces hmmsearch output files named by the HMM internal NAME field (via hits.query.name), but MacSyFinder --previous-run expects filenames matching the HMM filename stem. For DefenseFinder, 182/1178 profiles have mismatched internal NAMEs vs filenames, causing MacSyFinder to re-run hmmsearch from scratch (hours on large datasets).

Fixes applied in local Astra (to port to Aksha)

1. write_macsyfinder_hit() filename mapping (NEW)

• parse_hmms() now builds a {internal_name → filename_stem} mapping dict
• This mapping is threaded through hmmsearch() to write_macsyfinder_hit()
• Output files use the filename stem instead of the internal NAME
• Safe fallback: if no mapping exists for a name, falls back to internal NAME

2. Previously applied fixes (already in local ~/astra/astra/search.py)

• _write_macsyfinder_conf: db_type = ordered_replicon (was unordered)
• write_macsyfinder_hit: open("a") with header-once guard (was open("w"), overwrote per-FASTA)
• finalize_macsyfinder_files(): appends // end markers after all writes
• main(): cleans hmmer_results/ before new run to avoid stale appends
• initialize.py: GitHub API pagination (?per_page=1000 + Link header loop)

Why this matters

Without these fixes, MacSyFinder --previous-run silently falls back to running hmmsearch itself, which is orders of magnitude slower than Astra/PyHMMER. On DPANN (3.5M proteins, 1178 HMMs), this turns a ~5 minute job into a multi-hour job.

Affected databases

• DefenseFinder (182/1178 mismatched)
• TXSScan (not yet audited — likely similar issues)
• Any future HMM set where internal NAME ≠ filename

[jwestrob]

Additional fix: HMM chunking for --cascade mode (OOM on large databases)

Root cause: --cascade groups HMMs by cutoff type. For KOFAM (26,833 profiles), nearly all have GA thresholds, so they land in a single group. pyhmmer.hmmsearch(hmm_group, sequences, cpus=N) then loads all 26K HMMs into memory across N worker threads simultaneously → OOM.

Symptom: Astra parses all proteins and HMMs successfully, then silently exits with code 0 before the per-genome search loop starts. No error message, no output files.

Fix applied in ~/astra/astra/search.py:
Chunk hmm_group into batches of 500 before passing to pyhmmer.hmmsearch(). Same results, bounded memory.

HMM_CHUNK_SIZE = 500
for chunk_start in range(0, len(hmm_group), HMM_CHUNK_SIZE):
    hmm_chunk = hmm_group[chunk_start:chunk_start + HMM_CHUNK_SIZE]
    for hits in pyhmmer.hmmsearch(hmm_chunk, sequences, cpus=threads, **kwargs):
        ...

This is the 6th local Astra fix to port.