refactor: derivation.derive() pure function for ingest pipeline (#261)

[robotrocketscience]

Summary

Closes #261 (partial — see Scope below). Adds src/aelfrice/derivation.py exposing a pure derive(DerivationInput) -> DerivationOutput callable that takes raw text + source metadata and returns a Belief + Edge[] (or a skip with reason). Zero MemoryStore dependency — same input, same output, no I/O.

The derivation half of the pipeline is now testable in isolation and provides the pure callable the v2.x replay_full_equality flip-readiness probe needs. Each call site that goes through derive() is rewritten to: build input → derive → check duplicate → record_ingest + insert.

Changes

• New src/aelfrice/derivation.py (213 lines): DerivationInput, DerivationOutput, derive(). Two dispatch paths: classify-based (filesystem/git/python_ast/feedback_loop_synthesis/legacy_unknown) and lock-based (mcp_remember/cli_remember).
• New tests/test_derivation.py (319 lines, 23 tests): purity, each source_kind, lock priors, classifier rejection, unknown-kind ValueError.
• Refactored src/aelfrice/ingest.py and src/aelfrice/scanner.py (regex path) to call derive().

Scope: 2 of 6 entry points refactored

The issue body asked for all six entry points. Three were left direct-construction; reasons:

1. mcp_server.tool_lock and cli._cmd_lock — use sha256("lock\x00" + content) as belief id, a deliberately distinct namespace from derive()'s sha256(source_path \x00 text) scheme. Routing through derive() would compute different ids and silently break belief lookup for existing locked beliefs in production stores. Data-identity break, not a refactor — rejected.
2. triple_extractor._resolve_or_create_belief — uses sha256("triple\x00normalized_phrase") id namespace + fixed (1.0, 1.0) priors for noun-phrase entity beliefs. Same id-namespace incompatibility — rejected.
3. classification.accept_classifications — applies host-LLM classification results that have already been computed. Calling derive() here would re-run the regex classifier and potentially override the host's typed results. Behavior change, not a refactor — rejected.
4. Scanner's opt-in llm_router path — returns custom origin (e.g. DOCUMENT_RECENT) not representable in current DerivationInput. The default no-LLM regex path (production default) is fully routed.

These four sites need follow-up design work before they can be unified — id-namespace migration for #1/#2, classifier-replay handling for #3, DerivationInput extension for #4. Each is queen-tier, not mechanical refactor. Recommend filing as follow-up issues if the unification is wanted; otherwise leaving them as their own derivation paths is defensible.

Test plan

• tests/test_derivation.py: 23 new tests, all pass.
• Full suite: 1802 passed, 6 skipped, 1 pre-existing failure deselected (test_serve_raises_clear_error_when_fastmcp_missing — environment-dependent, fails on main HEAD too).
• Rebased onto current main (post fix: stamp beliefs.last_retrieved_at on hook-driven retrieval (#222) #266 merge); no conflicts.

Commits

• f258fa1 — add derivation.derive() pure function
• a76fb18 — route ingest through derive()
• 14ac869 — route scanner regex path through derive()

Out of scope

• replay_full_equality consumption ([v2.x] Implement replay_full_equality — flip-readiness probe #262).
• legacy_unknown migration ([v2.x] legacy_unknown migration for pre-v2.0 stores #263).
• The 4 unrouted entry points above — non-refactor work, separate issues if wanted.

[robotrocketscience]

Superseded by #273 (merged). Same target: extract pure derive() for #261.