feat: improve deterministic prose memory ranking

[suhaanthayyil]

https://entire.io/gh/entireio/entire-graph/trails/38

Summary

• add bounded, deterministic prose-only word-family matching for safe verb and derivational variants
• preserve the existing exact/plural path and keep code search behavior unchanged
• reserve more result tail for independent prose parents on list-style questions so evidence can span sessions
• document the fairness constraints and verification plan for this follow-up

This is a stacked follow-up to #84 and should not be merged before that PR.

Fairness and scope

• no benchmark-specific vocabulary, case IDs, expected answers, prompts, or holdout data in product code
• no reader, grader, API, credential, network, dependency, or provider change
• changes are gated to the existing Markdown/prose-parent path
• no paid QA benchmark was run for this branch
• the local retrieval audit used only a previously retired 30-case LOCOMO / 5-case LongMemEval-S tune selector; it is development evidence, not a publishable benchmark win
• a fresh blinded evaluation remains separately gated

Verification

• go test ./internal/sem -count=1
• go vet ./...
• go test -count=1 ./...
• go build -o /tmp/entire-graph-v53-ranking ./cmd/entire-graph
• candidate binary SHA-256: 06dac86768418bc44aec1134c695736db8449489d90e1eb83b1c5deb212466e7

Exact base-vs-new no-API retrieval audit on the retired selector:

Dataset	Cases	Base recall@10	New recall@10	Delta
LOCOMO	30	0.9388888889	0.9388888889	0.0000000000
LongMemEval-S	5	1.0000000000	1.0000000000	0.0000000000

All 35 per-case recall@10 values were identical; nonzero-to-zero regressions: 0.

Review

• task-level independent reviews: approved
• full-branch independent review: approved after three adversarial fix/re-review rounds with no remaining Critical, Important, or Minor findings