Batch embed_text trigram hashing (37-78% of parse_bundle time)

[valiantone]

Summary

Batch the per-trigram hashing in embed_text — it dominates bundle parse time.

Measured impact (spike #48, PR #84, B3)

embed_text is 37–78% of parse_bundle wall time across bundle trees (100–2500 bundles). It issues one hashlib.md5 call per character trigram (src/hotmem/embed.py:44). A native markdown/parse helper would optimize the wrong 20%; the lever is here.

Proposal

Batch trigram hashing (single hashlib update over a packed buffer, or a cheaper rolling hash), keeping hotmem-hash-v1 embedding semantics bit-compatible — otherwise version the embedding model string.

Acceptance criteria

• hotmem-hash-v1 vectors unchanged (or model string versioned + migration noted).
• Measurable parse_bundle speedup on the spike's bundle trees (fixtures reusable).
• No behavior change for callers.

Context

Spike recommendation follow-up #5: bench/native_spike/README.md. Related: #48, #49.

[valiantone]

Atlas alignment context — 2026-08-28

This issue is aligned as a contained public HotMem parse/snapshot optimization. HotMem remains the local-first portable memory foundation; EMOS owns enterprise reasoning, policy, orchestration, and promotion, while Memoria consumes HotMem-backed operational memory in its v0.2 demo world.

Decision boundary: preserve hotmem-hash-v1 vectors exactly. If any optimization changes vector output, treat it as a versioned embedding-model change with explicit migration and compatibility documentation—not as an invisible performance refactor. Snapshot, hydration, retrieval, and provenance consumers must retain their current observable behavior.

The benchmark measures parse_bundle performance only. It does not support a system-wide latency/SLO claim or authorize native dependencies, a persistent graph, enterprise control-plane behavior, or automatic feedback promotion.

Source: Atlas bounded-loop synthesis and HotMem native-helper spike #48.