fix(lineage): require exact calibrated active channel profiles

docs/adr/0064-lineage-evidence-and-tree-assembly.md

@@ -2,6 +2,8 @@
 
 - Status: Accepted
 - Date: 2026-08-19
+- Amended by: [ADR 0200](0200-channel-weight-reconciliation.md) and
+  [ADR 0205](0205-tepp-lineage-anchor.md)
 
 ## Context
 
@@ -16,9 +18,10 @@ or promoting an inferred relation to fact.
 - Treat every input record as a fallible mention and every accepted edge as a
   lineage instance supported by evidence, not as a proven business fact.
 - Fuse independent temporal, secondary-key, text/embedding, and optional LLM
-  channels through the RankWeave weighted convex fusion contract. A missing
-  channel is dropped and weights are renormalized; it is never replaced with a
-  fabricated negative or score.
+  channels through the RankWeave weighted convex fusion contract. Resolve a
+  missing channel before loading the exact calibrated active-channel vector;
+  never repair or renormalize a vector estimated for another channel set, and
+  never replace a missing channel with a fabricated negative or score.
 - Keep the channel-score breakdown and provenance on every candidate decision.
   Candidates below the minimum fused-score floor remain roots rather than being
   force-attached.

docs/adr/0172-event-lineage-channel-evidence.md

@@ -3,6 +3,8 @@
 **Status:** Accepted
 **Date:** 2026-08-21
 **Issue:** [#274](https://github.com/ContextualWisdomLab/LineageWeave/issues/274)
+**Amended by:** [ADR 0200](0200-channel-weight-reconciliation.md) and
+[ADR 0205](0205-tepp-lineage-anchor.md)
 
 ## Context
 
@@ -33,8 +35,9 @@ authoritative; PROV-O/RDF export is a projection.
    `lineage_signal_temporal`, `lineage_signal_secondary_key`,
    `lineage_signal_text`, `lineage_signal_llm`. The LLM row is omitted
    when the adjudication client is unavailable; it is never fabricated.
-3. Weights are the normalized active weights actually used
-   (`reconstruct.active_weights`). Contribution is `weight * score` and
+3. Weights are the exact calibrated active-channel vector validated by
+   `reconstruct.active_weights`; no different channel set is renormalized into
+   it. Contribution is `weight * score` and
    must reconcile with `fused_score` within
    `CHANNEL_EVIDENCE_TOLERANCE` (`1e-6`).
 4. Live Event Lineage is replaced atomically. A singleton
@@ -46,10 +49,10 @@ authoritative; PROV-O/RDF export is a projection.
    The administrator-triggered live rebuild and PostgreSQL import pass the
    configured contextual-orchestrator adjudication client through the same
    reconstruction boundary only when the exact candidate-pair count is at
-   most 5,000. Larger snapshots drop the LLM channel before any provider call
-   and renormalize the remaining weights. This is an operational work bound,
-   not a model-quality or provider-ranking heuristic. One rebuild never mixes
-   LLM and non-LLM weight profiles across edges.
+   most 5,000. Larger snapshots drop the LLM channel before loading the
+   separately calibrated three-channel profile. This is an operational work
+   bound, not a model-quality or provider-ranking heuristic. One rebuild never
+   mixes LLM and non-LLM weight profiles across edges.
 5. `GET /api/lineage` returns an additive `channel_evidence` collection
    on each visible edge (`signal_code`, `signal_label`, `score`,
    `weight`, `contribution`, `rank`) ordered by contribution, then
@@ -67,8 +70,9 @@ authoritative; PROV-O/RDF export is a projection.
   inference from source evidence.
 - A later rebuild rewrites live Event Lineage as a whole; historic
   meaning is not silently mutated in place.
-- Completeness is lower when the LLM channel is unavailable, matching
-  ADR 0064: missing channels are dropped and weights renormalize.
+- Completeness is lower when the LLM channel is unavailable; the exact
+  three-channel profile is loaded instead of renormalizing a four-channel
+  estimate.
 
 ## References
 

docs/product-requirements.md

@@ -31,7 +31,8 @@ invented facts, scores, weights, or identities.
 - Reconstruct bounded post-to-post parentage from independent channels.
 - Keep Event Lineage separate from ontology/KG edges.
 - Persist reconstruction profile and participating channel evidence.
-- Drop unavailable channels and fail closed on malformed calibrated weights.
+- Resolve unavailable channels before loading the exact calibrated active-set
+  vector; never renormalize a vector calibrated for a different channel set.
 - Never label an inferred lineage edge causal or authoritative without
   separate evidence.
 

lineageweave/embedding_client.py

@@ -1,8 +1,8 @@
 """Pluggable embedding channel.
 
 The default :class:`NullEmbeddingClient` makes the channel unavailable
-rather than faking a score -- ``reconstruct.active_weights`` drops and
-renormalizes around any channel whose client reports ``available = False``.
+rather than faking a vector; callers must use the calibrated profile for the
+channels they actually execute.
 :class:`ContextualOrchestratorEmbeddingClient` calls the authenticated
 contextual-orchestrator ``/v1/batch/embeddings`` boundary once a credential is
 set. No client in this repository calls a provider embedding endpoint directly.

lineageweave/lineage_persistence.py

@@ -65,8 +65,8 @@ def lineage_edge_specs(
 
     ``llm`` defaults to ``None``, which ``reconstruct()`` treats as the
     unavailable :class:`~lineageweave.adjudication_client.NullAdjudicationClient`
-    (the llm channel is then dropped and the rest renormalized, not
-    faked) -- callers that want the highest-weighted reasoning channel
+    (the supplied calibrated vector must omit the llm channel) -- callers that
+    want the reasoning channel
     actually contributing to real reconstructions must pass a real one.
 
     ``weights`` is required and always a psychometric estimate (ADR

lineageweave/reconstruct.py

@@ -11,6 +11,7 @@
 
 from __future__ import annotations
 
+import math
 from collections import defaultdict
 
 import rankweave as rw
@@ -25,8 +26,10 @@
 # today, TEPP when integrated) -- product paths load the persisted
 # corpus estimate, and the library demo estimates from its declared
 # generative design (channel_weight_estimation.estimate_fixture_channel_weights).
-# Every reconstruct() caller passes weights explicitly; the llm entry
-# renormalizes away when no client is configured (see active_weights()).
+# Every reconstruct() caller passes weights for the exact channel set it can
+# execute; a vector estimated for another set is never repaired or renormalized.
+
+_CORE_WEIGHT_CHANNELS = frozenset({"temporal", "secondary_key", "text"})
 
 # ponytail: only the most recent WINDOW prior records in a group are
 # considered as candidate parents, bounding per-group cost to O(n*window)
@@ -47,12 +50,23 @@
 def active_weights(
     llm: AdjudicationClient, weights: dict[str, float]
 ) -> dict[str, float]:
-    """Drop and renormalize the llm channel's weight when no client is configured."""
-    active = dict(weights)
-    if not getattr(llm, "available", False):
-        active.pop("llm", None)
-    total = sum(active.values())
-    return {channel: weight / total for channel, weight in active.items()}
+    """Validate and return the calibrated vector for the exact active channels."""
+
+    expected = set(_CORE_WEIGHT_CHANNELS)
+    if getattr(llm, "available", False):
+        expected.add("llm")
+    if set(weights) != expected:
+        raise ValueError("weights must exactly match the active lineage channels")
+    values = tuple(weights.values())
+    if any(
+        isinstance(value, bool)
+        or not isinstance(value, (int, float))
+        or not math.isfinite(float(value))
+        or float(value) <= 0.0
+        for value in values
+    ) or not math.isclose(sum(values), 1.0, abs_tol=1e-9):
+        raise ValueError("active lineage weights must be finite, positive, and sum to one")
+    return dict(weights)
 
 
 def _group_by(records: list[Record]) -> dict[str, list[Record]]:
@@ -159,10 +173,10 @@ def reconstruct(
         records: every record across every group; grouping happens here.
         llm: adjudication channel client; defaults to
             :class:`~lineageweave.adjudication_client.NullAdjudicationClient`
-            (the llm channel is then dropped, not faked).
-        weights: per-channel fusion weights before llm-availability
-            renormalization. Required, and always a psychometric
-            estimate (ADR 0145, second amendment): the persisted
+            (the supplied vector must therefore omit the llm channel).
+        weights: fusion weights calibrated for exactly the channels this call
+            executes. Required, and always a psychometric estimate (ADR 0145,
+            second amendment): the persisted
             fast-mlsirm corpus estimate on product paths, or
             :func:`~lineageweave.channel_weight_estimation.estimate_fixture_channel_weights`
             for the library demo. No hand-picked default exists.

tests/test_reconstruct.py

@@ -2,6 +2,8 @@
 
 from datetime import datetime
 
+import pytest
+
 from lineageweave import Record, reconstruct
 from lineageweave.fixtures import sample_records
 
@@ -80,6 +82,22 @@ def test_llm_channel_is_used_and_scored_when_a_client_is_supplied() -> None:
     assert all("llm" in edge.channel_scores for edge in tree_a.edges)
 
 
+def test_llm_weight_is_rejected_when_the_channel_is_unavailable() -> None:
+    """A four-channel estimate is not repaired into a three-channel vector."""
+
+    with pytest.raises(ValueError, match="exactly match"):
+        reconstruct(sample_records(), weights=_SYNTHETIC_WEIGHTS_WITH_LLM)
+
+
+def test_missing_llm_weight_is_rejected_when_the_channel_is_available() -> None:
+    """An active channel requires its own exact calibrated profile."""
+
+    with pytest.raises(ValueError, match="exactly match"):
+        reconstruct(
+            sample_records(), llm=_StubAdjudicationClient(), weights=_SYNTHETIC_WEIGHTS
+        )
+
+
 def test_candidate_window_bounds_which_priors_are_considered() -> None:
     records = [
         Record(f"r{i}", "G", f"record {i}", datetime(2026, 1, 1, i), "") for i in range(5)