R_eff: weakest-link cascade zeroes well-evidenced artifacts (ancestor zero overrides local evidence)

[explosivebit]

Summary

forgeplan's R_eff scoring propagates a zero from a weak ancestor down the
based_on / informs chain, overriding all local evidence on the downstream
artifact. The result: artifacts that are fully and correctly evidenced
(self_score = 1.0, multiple "Supports" evidence items each scoring 1.0) report
r_eff = 0, making them indistinguishable from genuinely unevidenced
artifacts. In a 309-artifact tree this produces ~52 false-zero artifacts and is
the main driver behind 185 weakest_link_unresolvable anomalies (>60% of the
tree).

Verified on forgeplan CLI/MCP (scores read via forgeplan_score).

Reproducer

# The cascade origin — a leaf design NOTE with NO evidence of its own.
forgeplan_score NOTE-003
#   self_score = 0.0
#   evidence   = []                      ← genuine zero (no local evidence)
#   r_eff      = 0.0
#   grade      = C

# Direct child — perfectly self-evidenced, yet zeroed by the ancestor.
forgeplan_score PRD-018
#   self_score = 1.0
#   evidence   = [EVID-027 (1.0, Supports),
#                 EVID-028 (1.0, Supports),
#                 EVID-033 (1.0, Supports)]
#   weakest_link = NOTE-003
#   r_eff      = 0.0                      ← WRONG: local evidence is perfect
#   grade      = B

# Two hops down — same story, zeroed transitively.
forgeplan_score PRD-021
#   self_score = 1.0
#   evidence   = [EVID-033 (1.0, Supports), EVID-034 (1.0, Supports)]
#   weakest_link = PRD-018
#   r_eff      = 0.0
#   grade      = B

Also verified with the same pattern (self_score 1.0, all evidence 1.0,
r_eff = 0 purely from an ancestor weakest_link): PRD-022 (3× 1.0
evidence, weakest_link = PRD-021) and PRD-023 (2× 1.0 evidence,
weakest_link = PRD-022). The chain is:

NOTE-003 (self 0, no evidence)
   └─> PRD-018 (self 1.0, 3× 1.0 evidence)  → r_eff 0
         └─> PRD-021 (self 1.0, 2× 1.0 evidence) → r_eff 0
               └─> PRD-022 (self 1.0, 3× 1.0 evidence) → r_eff 0
                     └─> PRD-023 (self 1.0, 2× 1.0 evidence) → r_eff 0

A single unevidenced leaf NOTE zeroes a 4-deep chain of otherwise-perfectly-
evidenced PRDs.

Expected behaviour

r_eff for an artifact should primarily reflect the quality of the evidence
that directly supports that artifact. An artifact with multiple
high-congruence "Supports" evidence items at 1.0 should not be reported as
r_eff = 0 solely because an ancestor it is based_on is itself weak.

Ancestor strength is a legitimate signal, but it belongs in a separate
metric (e.g. an ancestry/trust score), not as a multiplicative zero that
overrides locally-grounded evidence.

Actual behaviour

The weakest-link principle is applied across the entire ancestry chain, not
just within the artifact's own evidence set:
r_eff = min(local_evidence..., ancestor_r_eff). One unevidenced ancestor
zeroes the whole subtree regardless of how well each artifact in it is
evidenced.

Scale in a real project

• 309 artifacts, 185 weakest_link_unresolvable anomalies (>60%).
• ~52 of those are this bug: self_score = 1.0 + all local evidence 1.0,
  but r_eff = 0 via ancestor propagation (Class B).
• ~95 are genuine leaf-no-evidence zeros (EVIDs / unevidenced NOTEs) — a
  separate, known concern, see forgeplan_score returns r_eff=0 for leaf EVIDs with canonical bold-pattern bodies #325; do not fix that here.
• ~38 are detector depth-cap artifacts — filed separately (see the
  companion detector-imprecision issue).

The practical damage: forgeplan anomalies becomes un-triageable — a
maintainer sees 185 anomalies but cannot tell "genuinely unevidenced" from
"perfectly evidenced but poisoned by one weak ancestor" without manually
running forgeplan_score on every artifact.

Suggested fix direction

1. (Recommended) Decouple local vs ancestry scoring.
   • r_eff_local = min(local evidence scores) — weakest-link within the
     artifact's own evidence only.
   • r_eff_ancestry = min(ancestor r_eff) — kept as a separate, surfaced
     metric.
   • Report r_eff = r_eff_local as the artifact's primary score; show a
     composite trust separately if desired.
2. (Conservative) Propagate a zero only from a direct based_on parent to
   its immediate child, never transitively down the full subtree.
3. (Minimal floor) Ancestor propagation may not reduce r_eff below the
   artifact's own self_score when the artifact has valid local evidence.

Severity

Medium-high — not a data-corruption bug (the underlying evidence links are
intact), but it makes the headline quality signal (r_eff) and the anomaly
feed misleading for the majority of a mature artifact graph.

See also

• forgeplan_score returns r_eff=0 for leaf EVIDs with canonical bold-pattern bodies #325 — leaf-EVID/NOTE scoring (the ~95 genuine leaf zeros). Related
  but distinct; this issue is specifically about ancestor propagation zeroing
  well-evidenced artifacts.
• Companion issue: anomaly-detector imprecision (below-threshold reported as
  R_eff=0; depth-cap mislabelled as "cycle or depth cap").