e2e: cover semantic memory lifecycle and fallback

[codeaholicguy]

Current coverage

Lexical memory store/search is covered, but memory semantic status, download, reembed, hybrid search, and semantic write behavior are not.

What is missing

• Model-ready and model-missing status plus embedding counts
• Pinned download success, checksum rejection, partial failure, and offline behavior
• Reembed missing/stale rows, --force, counts, and idempotence
• Hybrid versus lexical retrieval mode and --explain
• Semantic-enabled store/update, embedding invalidation, and lexical fail-open behavior
• Project DB/config routing without downloading the production model in CI

Risk ranking

High, rank 5: recent releases focus on hybrid retrieval and noise reduction, but no E2E protects model packaging, configuration, or fallback.

Recommended E2E tests

• Use a preseeded tiny or fake model boundary in isolated HOME to assert status, download checksum rejection, reembed, hybrid mode, and lexical fallback.
• Initialize/install semantic-enabled config and assert MCP/CLI use the project DB, report missing embeddings, and preserve lexical availability.

Acceptance criteria

• The checklist tests are merged under e2e/.
• Tests are offline and deterministic; no production model download is required.
• JSON output asserts retrieval mode, readiness/counts, and fallback reason.

Related audit findings

A9: semantic store/update failures are invisible. C5: duplicate full-corpus count/read. D7: semantic scaling past the corpus cap.