feat(indexer): add opt-in raw-string substring ranking

docs/knowledge-operations.md

@@ -92,6 +92,20 @@ order regardless of the cue; `selection.ranking_applied` is false for those call
 Ordinary collection of other authorities and records remains available without
 the relevance opt-in. `KnowledgeStore.collect` delegates these same options.
 
+The default `relevance_method="weighted-lexical-v1"` searches the JSON
+representation of every field, including strings. Explicitly select
+`relevance_method="weighted-raw-string-v1"` (CLI:
+`--relevance-method weighted-raw-string-v1`) to match original string values
+instead: an actual newline or tab no longer creates a literal `n` or `t` match
+from its JSON escape. Literal backslash-letter text remains searchable.
+Nonstring fields, including the combined tags/applicability value, retain their
+legacy JSON representation. Both methods keep the same substring matching,
+distinct cue words, weights, positive-score admission and identity tie order.
+The nondefault method requires `relevance`; unsupported methods are rejected.
+Explicit IDs still bypass ranking. This option changes only owner discovery
+ranking, not query matching or optional context-resource scoring. It is a
+lexical-presence contract option, not evidence of general retrieval quality.
+
 `context_bytes` requests a separate context view for returned generated packaged
 skills. It requires an explicit positive integer budget, with no token estimate
 or default. It can be used without `relevance`; optional resources then match the

evidence/indexer-255-raw-string-20261030/README.md

@@ -0,0 +1,65 @@
+# Opt-in raw-string lexical contract
+
+Related tracker: [Issue #255](https://github.com/InferMatrixLab/InferRSI/issues/255).
+The single frozen comparison and independent saved-output audit passed **only the
+synthetic lexical-presence contract**. No general task accuracy, latency, model
+or hardware efficacy, default promotion or tracker completion is inferred.
+
+## Try the existing interface with the new explicit method
+
+From the source checkout with the package dependencies installed:
+
+```bash
+PYTHONPATH=src python -m inferrsi collect 'cpu.stat.local' \
+  --relevance cgroupv2 --relevance-method weighted-raw-string-v1 \
+  --context-bytes 65536 --limit 1 --json
+```
+
+Omit `--relevance-method` for unchanged legacy selection. The query remains a
+filter; the relevance cue ranks packaged skill owners. Context resources remain
+whole files with exact owner revision and SHA-256. The fixed real-corpus example
+returned the same complete 7,348-byte cgroup-resource-controls core in both
+modes; this is an interface demonstration, not a quality improvement claim.
+
+Synthetic cue `n t`, with unchanged title weight 4:
+
+| Original title | Legacy score | Opt-in score |
+|---|---:|---:|
+| Actual newline + tab | 8 | 0 |
+| Literal backslash-n + backslash-t | 8 | 8 |
+
+The phantom owner is no longer admitted. All declared genuine controls remain.
+See the [readable before/after](results/readable-demo.txt) and
+[independent audit](results/independent-postrun-review.json).
+
+## Retained evidence
+
+- Frozen baseline: `1052925ff22f9a850cebc64afd0083b35ed1fae9`.
+- Frozen candidate: `728984c524cf696d063ff7ef0688ae126c8546a6`.
+- [Exact protocol](plan/protocol.md), original [candidate patch](patch/frozen-candidate.diff)
+  and [Git bundle](patch/frozen-candidate.bundle). The Git bundle requires the
+  publicly retained baseline commit; it preserves the original local candidate
+  identity rather than substituting the later publication commit.
+- [Complete frozen study archive](results/frozen-study.tar.gz), SHA-256
+  `3074dbc24d198ff3d8ec440f839578340222b85600a5ff67c9d4186f88ec2b23`.
+  All 36 retained files are byte-preserved, including original fixtures,
+  prospective freezes, raw outputs, ordinary test logs, initial development
+  failure, superseded pre-review freeze, provenance, source/module fingerprints,
+  independent reviews and the one-run receipt. No scored comparison was repeated.
+- [Current-main integration check](results/integration-check.json): clean application
+  to `436c950cf06ded742fef61ff4303493ac5175d4b`, 108 affected tests passed,
+  2 skipped and Ruff passed. This does not rebase or replace the frozen study.
+
+## Results and boundaries
+
+All 14 scalar and 2 ordering fixtures matched their independently authored labels.
+Five phantom field-token contributions were removed across two scalar cases;
+15 original-string and 11 expected nonstring-JSON positive contributions were
+retained. Six nonstring field controls and 21 compatibility requests preserved
+legacy behavior, including optional context contents, hashes, revisions and
+omissions. Defaults, explicit-ID membership/order, filters, query and budgets
+are unchanged. Full ordinary CPU checks on the frozen candidate passed 2707 tests
+and 284 subtests with 12 skips; these are not task-efficacy evidence.
+
+See [publication provenance](plan/publication.md) for the distinction between
+new configured authorship, inherited historical evidence and CI qualification.