DR-150: A memoized extraction is keyed by the LLM identity that produced it

Context

Extraction memo entries keyed on (function fingerprint, content_text) only. The id-389 tier substrate runs the same extractors against three different answerers (mock sidecar, OpenRouter-served models, direct Anthropic), and the mock tier had memoized canned outputs under the same key a real-model run would hit — every tier-2/tier-3 “quality” run would silently replay mock outputs. Discovered S559 while executing AC-3 (the EXTRACTION_MODEL knob, owner-ruled for GLM-5.2 tier-2 extraction).

Decision

The effective LLM identity — resolved base URL (or anthropic-direct) plus model — is a required argument of every memoized extractor, so it participates in the memo key. Changing tier or model can never serve another identity’s memoized output. The signature change itself re-fingerprints the extractors: one whole-corpus re-extraction per DB on first deploy (owner-approved burn, $0 on mock).

Alternatives considered

• DR-060 manual version= bump per model change — burns the corpus on every swap, requires an operator to remember, and cannot fix cross-tier poisoning within one DB. Rejected.
• Sweep the memo store per tier switch — operational, unenforced, and destroys valid entries for the returning tier. Rejected.

Consequences

• Switching tiers back and forth is correct and cheap after the first walk per identity.
• OPEN (needs sign-off, id-389): ingest_file’s OUTER memo still short-circuits an env-only tier flip on unchanged bytes — the inner key protects changed bytes / full_reprocess / self-heal paths only. Closing it means adding the identity to ingest_file’s memo key, a change to the id-400-hardened surface.
• Sibling gap, unclosed: entity_pair_resolutions (pair_resolver’s DB cache) keys on (name_a, name_b, entity_type) — model-agnostic; needs a migration if the same rule is to hold there.

Implementation: canonical bf250ef28 (knob + identity threading + workflow tier wiring); grounding in tasks/id-389.md S559 entries.