fix(diff): stop destroying memories on CONFLICT and cosine-only UPDATE

[astrologist]

Problem

mnemon remember silently soft-deletes an existing memory whenever the built-in diff classifies the new content as CONFLICT or UPDATE. Both paths misfire on long technical notes, destroying unrelated memories:

1. classifySuggestion returns CONFLICT when similarity >= 0.7 and either text contains a negation word ("replaced", "no longer", "never", "deprecated", …). Change-log style memories almost always contain these words, and same-domain notes easily reach 0.7 similarity through shared vocabulary plus the cosine >= 0.85 hoist.
2. A CONFLICT is precisely the case where auto-deleting one side is wrong — the texts disagree, so the caller should decide which one wins.

We hit this three times in production use (v0.1.13): storing a new memory about topic A silently deleted an unrelated memory about topic B in the same entity cluster. Each time the only reason no data was lost permanently is that the replaced_id in the response was noticed in-session.

Reproduction

Store two same-domain notes describing different facts, both containing "no longer" (any two realistic incident write-ups about the same system will do). The second remember returns action: updated, diff_suggestion: CONFLICT and soft-deletes the first.

Fix

• remember: CONFLICT now keeps both memories. Detection is unchanged and the response still carries diff_suggestion: CONFLICT, so callers can merge or forget deliberately.
• remember: UPDATE only auto-replaces when the top match has TokenSimilarity >= 0.6. Cosine-only similarity (same-domain embeddings cluster at 0.85+) is not enough evidence to destroy a memory.
• classify: near-verbatim re-statements (token similarity > 0.9) are classified DUPLICATE before the negation scan — previously an exact duplicate containing a negation word was classified CONFLICT (masked by the old replace behavior; exposed once CONFLICT keeps both).
• classify: a new text >25% longer than its match is never DUPLICATE. Previously a superset extension (old text + a new sentence) was cosine-hoisted above 0.9, classified DUPLICATE, and skipped — silently dropping the new content.

Testing

• go test ./internal/search passes (call sites updated for the new tokenSim parameter).
• Manual scenario matrix in an isolated MNEMON_DATA_DIR (with Ollama embeddings active):

scenario	before	after
same-domain, different facts, negation words	CONFLICT → old memory deleted	CONFLICT → both kept
exact re-statement containing negation words	CONFLICT (replace)	DUPLICATE → skipped
superset extension (old + new sentence)	DUPLICATE → skipped, new info lost	UPDATE → replaced
unrelated new content	ADD	ADD

🤖 Generated with Claude Code