fix: report selector/index signature-grammar mismatch in efficacy

[karngyan]

Claim

autophagy mutations efficacy gave a misleading verdict when an installed mutation's trigger selector uses an older signature grammar than the index it is evaluated against.

The mutation mut_b13f2878…'s selector is failure/v1|shell|cd zuzoto && go build ./... 2>&1|exit:1, but the showcase database's event_signatures index was re-minted to grammar v2 by reindex --index-tool-input. The v1 operation key (operation/v1|…) then matches zero rows even though the 8 pre-install failures exist — they are simply indexed under v2. The report read 0 → 0 failures · no prior baseline · needs more observed occurrences, as if the failures never happened.

The honest statement is that the selector's grammar is older than the index's and the measurement cannot see those events. A v1 selector is not silently translated to v2 — the v2 normalization is not derivable from the v1 string.

Change

• New insufficient_reasons variant selector_grammar_mismatch (efficacy result v0.1, additive enum member). It fires when a selector's grammar version is older than the newest grammar present in the index and the index holds no rows of that older grammar. A partial/skipped reindex that still carries the old grammar does not trip it (the selector can still match those rows).
• autophagy-efficacy: EfficacyObservations gains index_grammar_versions; evaluate parses each selector's grammar version and compares.
• autophagy-store: new read-only Store::signature_grammar_versions() (no migration).
• CLI: renders the actionable sentence selector uses signature grammar v1 but the index is v2 — re-propose from current findings to measure efficacy, suppressing the consequential "needs …" reasons.
• Docs: result.schema.json enum value, a real valid/selector_grammar_mismatch.json fixture (generated from the reindexed showcase DB), spec README, and an addendum to ADR 0015 justifying why an additive enum member is not a version bump.

Evidence

Real-data verification on the reindexed showcase database (AUTOPHAGY_CONFIG_DIR=$(mktemp -d)), --now 2026-11-16T00:00:00Z. The index holds only v2 signatures (34,911 rows, zero v1); the installed mutation's selector is v1.

Before

insufficient data · 0 → 0 failures (0.0/wk → 0.0/wk) · no prior baseline · 80.5% classifiable · needs more observed occurrences

After

insufficient data · 0 → 0 failures (0.0/wk → 0.0/wk) · no prior baseline · 80.5% classifiable · selector uses signature grammar v1 but the index is v2 — re-propose from current findings to measure efficacy

insufficient_reasons in the JSON report is now ["selector_grammar_mismatch", "sparse_occurrences"].

New unit tests: v1 selector + v2 index → mismatch reason; matching v2 grammars unaffected; an old selector still present in a mixed index is not flagged; store grammar-version enumeration. mise run check is fully green.

Privacy

None. The report continues to cite exact event_ids and counts only — never command text or event content. The new store query reads grammar-version tokens from the existing redaction-approved event_signatures projection; nothing new is persisted and nothing leaves the machine.

🤖 Generated with Claude Code