SKT-0021.01: 118 instrument_mismatch gaps are all reality=unknown: decide capture enrichment or absent-evidence

Description

All 118 instrument_mismatch coverage gaps have only-in-reality=[unknown]. Verified 2026-08-29: not one of them is a case where synthkit publishes the wrong instrument type. Every one is the corpus failing to record an instrument type at all, so the comparator compares a real synth verdict against an absence and reports a mismatch.

This is the largest single block of noise in the report, it can never be cleared by changing synthkit, and it will still be there after every emission task in this epic is done.

Two honest routes, and the task should pick one with its reason rather than split the difference:

Enrich the capture so an instrument type is recorded where it is derivable. Note the existing derivation at e2e/receiver/receiver.go:234 infers a classic histogram from the _bucket/_count/_sum name suffix rather than from an observed bucket series — SKT-0010.07 already stopped work over exactly that weakness, so any enrichment must not repeat it. A suffix is not an observation.
Treat unknown as absent evidence and stop emitting a finding for it, consistent with SKT-0010.01’s decided rule that absent evidence is a gap rather than a divergence — here, arguably not even a gap.

Route 2 is cheap and defensible; route 1 is better evidence. What must not happen is a filter that hides the finding without recording that the corpus does not know.

Acceptance Criteria

#1 The chosen route is recorded with its reason, not split between both
#2 If the corpus is enriched, the instrument type comes from an observed series and never from a name suffix
#3 If unknown is treated as absent evidence, the report still makes clear the corpus does not know, rather than silently dropping the family
#4 The instrument_mismatch class is measurably reduced and the movement is attributed

Definition of Done

#1 make signal-fidelity runs and the instrument_mismatch count is reported before and after

Implementation Plan

Lane F chooses one honest unknown-instrument evidence policy, keeps unknown visible, and measures attributed comparator movement.

Implementation Notes

2026-08-30 closeout: chose absent-evidence handling because unknown is not an observed instrument type. The 118 former instrument_mismatch rows moved to the explicit unknown_instrument_evidence coverage class; instrument_mismatch moved from 118 to 0 without hiding corpus uncertainty. Final signal-fidelity gate exited 0.

2026-09-06 disposition, case (b): AC2 is conditional on corpus enrichment, which was not the chosen route. The recorded decision moved unknown types into explicit absent-evidence findings. Task stays Done with AC2 open as not applicable.

Final Summary

Unknown corpus instrument types remain visible as attributed absent evidence and no longer masquerade as synth shape mismatches.