fix: provision embedding model in lifecycle + stop rendering RRF scores as percentages (#240)

[edheltzel]

Fixes the root cause of #240 and the score-display bug found alongside it.

Found while diagnosing a 23-day silent semantic-search outage on a live install: embeddings = 0 rows, hybrid search degraded to keyword-only, and a natural-language query returned No results found for a memory that was sitting in L1 context at every session start.

1. Model provisioning — the root cause (#240)

The default embedding model swapped nomic-embed-text/768 → qwen3-embedding:0.6b/1024 in a72fac2 (#107/#160, 2026-06-22). No lifecycle script ever pulled it.

$ rg "ollama pull|qwen3|embedding" install.sh update.sh lib/install-lib.sh
(nothing)

Every install that updated past #160 without manually running ollama pull silently lost semantic search. Not a local misconfiguration — a migration that shipped without a migration step.

Fix: recall_provision_embedding_model() in lib/install-lib.sh, called from install.sh (self-check) and update.sh (step_verify) — update.sh being the path that actually regressed.

Two constraints shaped it:

• The model name is read from src/lib/embeddings.ts, never hardcoded in bash. A duplicated constant is exactly how the last swap drifted; hardcoding here would rebuild this bug for the next swap. Verified: no model name appears in any lifecycle script outside a historical comment.
• Respects the 🔒 Ed-locked OPTIONAL invariant (src/lib/embedding-marker.ts:10-13). No Ollama → informational, exit 0, never gates the install. Both paths exercised:

# ollama up + model present
✓ Embedding model present: qwen3-embedding:0.6b            exit=0

# ollama absent
→ Ollama not installed — semantic search stays disabled (optional; keyword search is unaffected)
→   To enable later: install Ollama, then 'ollama pull qwen3-embedding:0.6b' and 'recall embed backfill'
                                                            exit=0

2. The score field was polymorphic

score carried raw FTS5 bm25 rank on the FTS-only path (negative, unbounded) and RRF on the fused path (0..~0.033). One formatter rendered both as (score * 100).toFixed(1) + '%'.

Observed output: -1121.0% while degraded, then a meaningless 1.6% against a 3.3% ceiling once semantic recovered.

• mcp-server.ts — run the FTS-only list through reciprocalRankFusion so score has one unit everywhere. Same ordering, honest unit.
• search-fallback.ts — render rrf=0.0328, not 3.3%. RRF is a relative ranking signal, not a confidence; the % implied a meaning it never had.
• mcp-server.ts reused formatHybridResults instead of its hand-copied duplicate — the bug existed in two places and had to be fixed twice. One definition now (DRY law).

Tests: fixture score: 0.83 → 0.0328 (0.83 was never a value RRF could produce — the fixture encoded the misconception). New regression test asserts no % is ever rendered and a negative score cannot print -1121.0.

Verification

End-to-end on a live install: model pulled → 14,616 embeddings backfilled → vec index rebuilt → semantic backend: knn. The query that returned No results found now returns the decision it was looking for.

	before	after
MCP hybrid	keyword-only: embeddings unavailable; semantic backend: none	hybrid: FTS5 + embeddings; semantic backend: knn
embeddings / vec index	0 / 0	14,616 / 14,616

bun run lint clean. Full suite: 1236 pass.

On the 10 failures: all are recall_configure_opencode_mcp / recall_configure_pi_mcp and are a pre-existing worktree artifact, not from this PR. Proven by stashing every change and re-running on pristine main code inside the same worktree — identical 5 failures per file. Clean main passes 1245/0. Worth its own issue: AGENTS.md mandates workers run in a worktree, so every worker currently sees these 10 spurious failures.

Not included

• recall repair --execute leaves the vec index empty — semantic search silently degrades to bruteforce, and doctor reports PASS #241 (repair --execute leaves the vec index empty) — filed separately, needs its own fix.
• Item 3 / feat: query intent classification for hybrid search — auto-derive --bias-type + signal weighting from the query (extends #82) #234 (terse queries miss the right memory) — reproduction and a new RRF k=60 flattening finding added as a comment on feat: query intent classification for hybrid search — auto-derive --bias-type + signal weighting from the query (extends #82) #234. Deliberately not fixed here: it needs a relevance harness, and feat: query intent classification for hybrid search — auto-derive --bias-type + signal weighting from the query (extends #82) #234 is triaged "STEAL LATER".

Refs #240, #241, #226, #234, #107, #160.