vdf: overlap-path descriptor peel likely drops a real stock on SimService/Base.vdf (c Identified Oil Reserve)

[bpowers]

Problem

On third_party/uib_sd/spring_2008/SimService_BUENO/SimService/Model Files/Base.vdf (and its ctxt0001 sibling), the OVERLAPPING-descriptor peel path in src/simlin-engine/src/vdf/record_results.rs::identify_descriptor_records (mirrored in tools/vdf_xray.py) classifies the record for c Identified Oil Reserve as a graphical-function descriptor and drops its series.

Details:

• Its span is OT[17..19): a 2-element stock-coded span with real data blocks, f11=17 with n_lookups=18.
• The span overlaps AGE SPECIFIC FERTILITY DISTRIBUTION FUNCTION (OT[18..100), an 82-wide stock-coded span).
• The lexical lookup-def-name test is ambiguous for the pair, so the highest-f10 fallback decides -- and it picks c Identified Oil Reserve as the descriptor.

This is very likely a false positive: the file's other c Identified * Reserve siblings (Coal / Gas / tar sands) are all real 2-element stocks that are correctly kept.

Why it matters

Correctness: a real stock's simulation series is silently dropped from decoded results, which corrupts any comparison or import based on this file.

Components affected

• src/simlin-engine/src/vdf/record_results.rs (identify_descriptor_records, the highest-f10 fallback)
• tools/vdf_xray.py (same function, kept in parity)
• docs/design/vdf.md -- "Name-to-OT mapping" step 2 and "Appendix: the owner/descriptor discriminator"

Context

Identified while hardening the STANDALONE lookup-only descriptor drop (GH #840 follow-up work on branch vdf-parity-fixes). The behavior is pre-existing: identical before and after that branch's changes. The standalone-drop hardening in that branch fixed the four STANDALONE false positives (Agriculture Employment, Alaska Oil Discovered/Undiscovered, Atmos UOcean Temp) but deliberately did not touch the overlap path.

Possible approaches

• Strengthen the overlap-path discriminator so a span with real data blocks and stock-coded siblings (e.g. matching c Identified * Reserve family members that are kept) is not peeled as a descriptor.
• Prefer structural evidence (data-block presence, span width vs n_lookups consistency) over the highest-f10 fallback when the lexical lookup-def-name test is ambiguous.

[bpowers]

Investigated as part of the vdf-parity-fixes branch (three review rounds). Confirmed the data loss is real: c Identified Oil Reserve is a genuine 2-element stock (OT [17,19), series 51000->8784 and 7900->2958, initials matching c Initial Identified Oil Reserve) that the overlap-path f[10]-highest fallback deterministically peels because it is the component's only f11 < n_lookups candidate against the 82-wide AGE SPECIFIC FERTILITY DISTRIBUTION FUNCTION span.

A principled fix was designed and then rejected by corpus evidence: requiring the fallback's chosen candidate to pass the standalone-drop physical + consumer-corroboration gates would exonerate this stock (its forward link fails both the width gate, word[11]=1 != 2, and corroboration, word[10]=445 with no span starting there), BUT the same gates fail on 9 of Ref.vdf's 16 legitimate fallback peels: six Specified * scalar descriptors whose forward links land at the starts of 7-wide shared consumers (corroboration-by-length fails), plus RS HFC4310mee (width 63 != 7, the shared-2D-consumer family), and two zambaqui Pop-4/5 peels also flip. Gating the fallback therefore resurrects 9 ghost columns on Ref.vdf and breaks the C-LEARN comparator pins.

Conclusion: overlap-peeled descriptors legitimately include scalar members of shared-consumer families, so the standalone gates do not transfer to the overlap path. Any future fix needs a new discriminating signal (or a model-guided reader). The standalone-drop path itself was hardened on that branch (consumer corroboration + per-file coherence veto + veto diagnostics), which fixes the sibling false-positive that dropped four SimService scalar stocks.