fix(recall): importance filters, it does not rank

[MXAntian]

The measurement

A self-rated field that does not predict use was moving results 2-4 places.

RRF separates adjacent ranks by ~0.0026. The importance term at weight 0.05 buys:

gap	score delta	≈ rank positions
imp 8 vs 7 (the common case)	0.0050	2
imp 9 vs 7	0.0100	4
imp 10 vs 3	0.0350	17

A top-8 injection is decided inside that range — four places is the difference between a row reaching the model and not.

What it was spending that on

On a live 8.6k-row store:

rows ever recalled    importance avg 7.70
rows never recalled   importance avg 7.36     Δ 0.34 on a 1-10 scale
90.1% of the corpus at >= 7

Self-rated, saturated, uncorrelated with use.

It had already been demoted once, from a heavy weight to 0.05, and the comment beside it already read "don't let self-rated importance drive ranking". 0.05 was still enough to drive it.

What changes

Removed from both scoring paths. FTS weights renormalized across the three surviving signals (0.61 / 0.22 / 0.17) so the score keeps its 0-1 shape.

importance keeps the jobs it is good at — min_importance filtering, surface-cold thresholds, display. Filtering on it is a caller stating a floor; ranking on it is the store guessing.

On the test

Calibrated to the failure regime rather than an obvious case. My first version gave the low-importance row a large relevance edge — which survived the tilt, so it passed with and without the fix and proved nothing.

It now ties relevance and gives the low-importance row a single prior access, worth ~0.0455, deliberately less than the ~0.07 an importance gap of 10-vs-3 was buying:

before fix:  ✗ evidence of use outranks a higher self-rating
             used(imp3,acc1)@1  rated(imp10,acc0)@0
after fix:   ✓

level-rank-offset.test.mjs pins the composite as source text, so its regex moved with the formula; its actual assertion (no level-weight multiplier) is unchanged.

325 passed / 0 failed.

🤖 Generated with Claude Code