Deferred Path Search 42 EN

Execute auxiliary path search only when selected — experiment 42

A new Go path genuinely omits auxiliary index construction and ranking when auxiliary search is declined. This differs from merely counting a precomputed result as unused. No useful selector has been trained here, and actual LLM latency/token savings remain unmeasured.

The plan was committed first. Existing Rank computes baseline, normalized and interleaved orders for comparison. New RankSelected ranks the baseline and passes only experiment 41's 16 numerical features, by value, to a selector. Auxiliary indexing/ranking happens only after a true result. A nil selector skips feature extraction too and returns baseline-only ranking.

Path	Result	Actual auxiliary work
Nil selector	Full baseline order	0 builds, 0 rankings
False result	Full baseline order	0 builds, 0 rankings
Selector error	Baseline with selector_error	0 builds, 0 rankings
True, successful auxiliary execution	Full interleaving preserving the baseline's first candidate	Build when needed, then rank
True, auxiliary failure	Baseline with auxiliary_error	Count actual failed attempts
Invalid baseline	Error	Selector is not called

Feature/interleaving errors also have explicit baseline fallbacks. Private selector error messages are not copied into the result. No candidates are removed and no answer is treated as certain; auxiliary results only reorder the same candidate set.

WorkStats counts actual attempted operations: baseline ranking, auxiliary index construction and auxiliary ranking, including failures. These are not model-call counts or estimates of time/memory/tokens saved. Future total-cost measurements must also include baseline index construction, features and the selector itself.

Shared baseline/auxiliary operations were factored out of the existing Ranker. Old ranking results are checked against an independent pre-refactor reference. The standalone function retains no index between calls. Only Ranker.RankSelected opts into exact-full-catalog index reuse. Experiment 35 did not establish a real-workload benefit from reuse, so it is not activated by default.

A synthetic reused-worker sequence skip → use → use → skip → change catalog and use records five baseline rankings, two auxiliary builds and three auxiliary rankings. After the first skip, no auxiliary index exists. Skipping with a warm auxiliary index still performs zero auxiliary rankings. Reused objects may retain the previous index memory; skipping does not imply freeing that memory.

Tests cover always-on/off and selector-error policies, lazy first construction, reuse and catalog mutation, independent prior results/feature arrays, long queries/10,000 candidates and independent workers under race detection. Punctuation-only queries normalize to empty text and fall back in the existing reference too; the new path preserves this. Queries that exceed 128KiB only after normalization also retain the full baseline order. Full race/vet/format/redacted scans gate publication.

One sequential worker owns each Ranker. Selectors must not mutate input paths or re-enter the same object. Independent workers use separate objects and immutable models; no global cache or lock is added.

This remains an internal research API and does not automatically connect learned weights to existing CLI commands. Next, fit a selector under fixed development/source-eligibility rules and compare baseline-only, always-auxiliary and learned selection using actual computation and quality. These synthetic execution checks are not a new 2,400-task user evaluation or a training success. With acquisition running separately, they are not latency/memory benchmarks. No final evaluation or new HF release occurs here.

Source / 원본 · PR #65