research-161-r1: QualifyArtifact — PROCEED verdict with source-verified findings

research/issue-161/qualify.json

@@ -0,0 +1,16 @@
+{
+  "stage": "QUALIFY",
+  "issue": 161,
+  "round": 1,
+  "outcome": "PROCEED",
+  "findings": [
+    "INTERFACES UNAMBIGUOUS (verified against source): every line anchor cited in the build instructions exists in the current tree — fingerprint() at src/rune/orchestration/benchmark.py:22; MemorySelector.add_source(name, fn) at memory_selector.py:130 with the exact 4-arg call contract (query, types, MAX_CANDIDATES, min_score) at memory_selector.py:160; DEFAULT_SOURCES registry at memory_selector.py:504-511; RECALL_BUDGET=8 at memory_store.py:54; result cache at memory_store.py:605-617; _embed content-hash cache at 333; probe-limit RECALL_BUDGET*2 at memory_store.py:639 and 689; the three mutually exclusive fusion branches at 731/737/741; metadata composite multipliers at 796-800 with datetime.now() recency at 781-789. A coder can implement without guessing at any cited location.",
+    "The recall() extension spec is concrete: exact new signature given (recall(db, query, budget=RECALL_BUDGET, min_score=RECALL_MIN_SCORE, *, probe_cap: int = RECALL_BUDGET * 2)) matching the existing signature at memory_store.py:575; 'mode' field values enumerated (lexical-only | dense-only | hybrid) and mapped one-to-one onto the three verified fusion branches; default-preserving requirement is backed by an explicit regression test (test_recall_cap_parameter_defaults_preserve_behavior).",
+    "IMPLEMENTATION SEQUENCE IS ORDERED BY REAL DEPENDENCIES: M1 (mode field) precedes the harness because mode assertion (INV-1) requires the field to exist; M8 (probe_cap) precedes the sweep steps; M6 isolation precedes any write; M10 shared schema precedes harness record emission; backends-under-test and scale sweep come after the harness core exists. No step depends on a later step; each step names its files and its tests.",
+    "ACCEPTANCE CRITERIA ARE OBSERVABLE: AC-1 through AC-12 are each evaluable by an automated test (bit-identity across service restart, zero records on unreachable embedder, PR-curve shape in output, delta-column presence, p99 rows at 1K/10K/100K, non-clipped k=32 sweep, caveat/drop verdict on graph row, assertion-log proof of isolation, headline-metric presence, per-stage timing presence, single-schema check, authority-map diff empty). tests_required lists 16 named tests that map onto the ACs and invariants with no gaps in coverage of M1-M10.",
+    "AUTHORITY BOUNDARIES RESPECTED: the harness is evaluation-only and leaf-positioned; it registers backends through the EXISTING add_source/DEFAULT_SOURCES contract (no second protocol — the memory-adapter-protocol alternative is explicitly rejected with rationale); memory_store.py touches are limited to an observability field plus a default-preserving parameter; config/v1_authority_map.json is unchanged; INV-13 explicitly forbids consuming the mode field for behavior. No step contradicts the stated boundaries.",
+    "ROLLBACK IS REAL: deleting the new harness file, its tests, and harness-only registry registrations removes the capability; the three production touches (mode field, probe_cap param, inert schema addition) are independently revertible and guarded by the default-preserving regression test; the fallback state (research-reasoned constants, no quality measurement) is named.",
+    "MINOR NON-BLOCKING NOTES for the implementing coder (resolvable without re-research, do not change any interface): (a) the mechanism for reading the served embedding-model digest from Ollama (e.g. /api/show or blob sha256) is not spelled out — any reasonable discovery method satisfies the stated pin-and-verify contract; (b) the pinned-dataset file's internal record schema (query/graded-relevance/task-class fields) is left to the harness implementation — acceptable because the dataset lives entirely inside the new self-contained harness file and its tests, and only its SHA-256 pinning is contract-relevant; (c) test_evidence_citation_gate describes a 'decision-input lint' whose exact placement (test vs lint script) is flexible — either satisfies FT7/INV-7."
+  ],
+  "reopen": null
+}