docs(evidence): finalize ticketed restructuring specification

docs/adr/0001-evidence-publication-boundary.md

@@ -0,0 +1,7 @@
+# Use workspace activation as the Evidence publication boundary
+
+Curated Evidence becomes Published Evidence only when it is valid, belongs to the
+active workspace revision, and belongs to the atomically published Evidence generation.
+Human Git review remains required before activation, but its approval is not duplicated
+as mutable state in the Evidence manifest; this keeps Git and the workspace registry as
+the existing sources of truth instead of introducing a second approval mechanism.

docs/adr/0002-kind-independent-evidence-identifiers.md

@@ -0,0 +1,8 @@
+# Keep Evidence identifiers independent from kind
+
+An Evidence Unit keeps the same identifier when its kind is corrected or its Source
+Evidence is unambiguously renamed; kind remains a separate validated field. This avoids
+breaking citations and evaluation fixtures for a classification change, while genuine
+semantic splits receive new identifiers so independent units never share an identity.
+Identifiers use `evidence:<slug>`, are assigned once, persisted in the manifest and are
+never recomputed automatically from mutable titles, paths or content hashes.

docs/adr/0003-evaluate-evidence-before-activation.md

@@ -0,0 +1,7 @@
+# Evaluate candidate Evidence before activation
+
+Evidence preprocessing builds a candidate vector generation and runs the workspace's
+versioned evaluation set against that exact generation before atomically activating it.
+The workspace revision may temporarily have no matching active Evidence during this
+maintenance window, which is accepted as fail-closed degradation instead of introducing
+a distributed transaction across Git, the workspace registry and Qdrant.

docs/adr/0004-formula-evidence-is-an-expression.md

@@ -0,0 +1,6 @@
+# Treat Formula Evidence as a composable SQL expression
+
+Formula Evidence contains one validated PostgreSQL expression with declared input
+columns, not a complete query or statement. This makes formulas safely composable by
+SQL generation; complete documented queries remain Example Evidence, and incompatible
+legacy formulas require human review during migration.

docs/adr/0005-evidence-contributes-to-semantic-stages.md

@@ -0,0 +1,13 @@
+# Evidence contributes to semantic workflow stages
+
+The Evidence Module is a contributor to the existing semantic stages, not a visible
+workflow stage. `clarification`, `rewriting`, `schema_linking`, `cte` and `final_sql`
+query it with their corresponding purpose; `memory` remains owned by the Memory Module
+and `synthesis` performs no Evidence search. Each mapped stage searches independently
+and the workflow records only a minimal receipt containing stage, purpose, vector
+generation and returned Evidence IDs.
+
+A successful search may return no matches and does not block the stage. Technical
+unavailability is a distinct typed outcome that blocks the calling stage until retry,
+without stale-generation or purpose fallback. This favors an explicit temporary stop
+over silently treating a broken Evidence dependency as absence of domain knowledge.

docs/adr/0006-grounded-atomic-evidence-preparation.md

@@ -0,0 +1,12 @@
+# Ground and atomically apply Evidence preparation
+
+Every proposed Evidence Unit carries one to five short supporting excerpts that can be
+found in its normalized Source Evidence. The model may reuse only identifiers supplied
+for previous units; deterministic application code assigns all new canonical IDs. This
+makes provenance and identity mechanically checkable without pretending that automated
+validation can replace human semantic review.
+
+Preparation validates the complete changed batch in a temporary area and applies
+curated files plus the manifest atomically. A model timeout or invalid response is not
+retried automatically and leaves the worktree unchanged. Explicit `evidence resolve`
+actions retire or relink units while preserving an ordinary, recoverable Git diff.

docs/adr/0007-add-bm25-without-rebuilding-qdrant.md

@@ -0,0 +1,13 @@
+# Add BM25 without rebuilding the shared Qdrant collection
+
+The workspace keeps its existing unnamed dense vector and adds only the sparse `bm25`
+vector with IDF through Qdrant's additive vector-schema operation. Only Evidence points
+are repopulated with both default dense and BM25 values. Schema, Memory and solved
+questions retain their current dense points and are verified before and after the
+upgrade.
+
+This replaces the planned destructive conversion to a named `dense` vector. If BM25 is
+missing, Evidence preprocessing may add it and verify the resulting schema; session
+runtime remains read-only. An incompatible existing BM25 definition fails without
+mutation. A candidate failure leaves the additive schema in place while Evidence stays
+unavailable, avoiding data loss in the other workflow modules.

docs/adr/0008-make-hybrid-evidence-retrieval-deterministic.md

@@ -0,0 +1,24 @@
+# Make hybrid Evidence retrieval deterministic and diagnosable
+
+Dense and BM25 retrieval receive the same deterministic query text. It preserves the
+original question and appends nonempty concepts, tables and columns in a fixed order;
+question and context receive Unicode NFC, newline canonicalization and outer trimming.
+Context values are then deduplicated exactly and sorted, without lowercasing. Case,
+punctuation and internal whitespace remain intact. This avoids accidental ranking
+changes caused only by metadata ordering, preserves quoted PostgreSQL identifiers and
+keeps the two retrieval branches directly comparable.
+
+Evidence fragmentation follows semantic headings, typed fields and paragraph
+boundaries. Formulas, value/meaning pairs, mappings, rules and URLs remain atomic. An
+atomic element larger than the existing `max_chunk_chars` limit creates the blocking
+`atomic_content_too_large` Review item rather than being split mechanically. The limit
+applies to the complete rendered text, defaults to 4,000 characters and is not
+duplicated by an Evidence-specific setting.
+
+An L0 contract test starts the exact Qdrant image referenced by `compose.yaml` and
+proves Italian server-side `qdrant/bm25` ingestion and search in a temporary collection.
+There is no FastEmbed or dense fallback for Evidence when that capability is absent.
+
+The versioned evaluation set contains lexical, semantic and mixed queries. Its report
+shows dense-only, BM25-only and fused ranks for expected Evidence. Publication remains
+governed only by the simple fused top-10 rule; branch ranks and hit@5 are diagnostic.