Streaming-subset completeness audit: every lowering rejection is now a product gap

[FBumann]

With the streaming engine as the only runtime lane (#26), every RelationalBuildError raised by lowering is a product limitation, not a routing event. Audit each rejection site and decide: keep as a documented language boundary, or implement.

Known list (grep raise RelationalBuildError in lowering.py):

• ** operator — likely stays out (nonlinear); error message should suggest the epigraph/piecewise alternatives.
• Custom Python helpers — stays out by design; the error already points to macros:.
• where-comparisons on dimensions ("snapshot > 2") — implement: the dim ordinal is in the executor's dim tables; SPEC §6 documents the semantics; the eager oracle supports it. Probably the highest-value gap (time-windowed constraints).
• Sum over absent dims of constant parts — keep (ambiguous under masks), but check the message quality.
• Verify Backend router, vtype, shift(), and piecewise: blocks #19's vtype/piecewise:/shift execute end-to-end on the streaming engine at scale, not just at test size.

Definition of done: for each site either a differential test proving support, or a documented boundary with an actionable error message. ARCHITECTURE.md's language section updated accordingly.

🤖 Generated with Claude Code

[FBumann]

Oracle sunset trigger — recording the decision point this audit will make
answerable, so it gets revisited deliberately rather than drifting.

The compat lane exists for two jobs: Python math the language cannot say
(#38), and being the differential oracle. The second one is what costs:

• ~870 package lines (builder.py 268, loader.py 225, compat.py 174, the
  eager helpers.py bodies 112, where_parser.evaluate_where 90)
• 84% of the test suite — 2,644 of 3,130 lines, because most feature tests
  are differential
• a two-implementation tax on every new primitive (ARCHITECTURE, "taxed
  primitives")

It is worth paying while the core language is still moving, and its value is
that it is external: replacing it with golden LP fixtures would only check
the executor against itself, so a consistently-wrong executor would stay wrong.

Two forces erode it independently of any decision:

• coverage shrinks as the language diverges — escape: islands are outside
  the harness by construction, and sink-bounded features (quadratic, SOS piecewise: method 'sos2' (native SOS stream) and 'lp' (tangent lines, zero aux vars) #23)
  cannot be compared at all
• each Track 1 primitive pays the tax up front

Proposed trigger: revisit once Track 1 items 1-6 have landed and this
audit produces a coverage number. At that point the question is concrete —
"we pay ~870 + 2,644 lines for independent verification of a language that no
longer changes" — and can be answered with data instead of instinct. Until
then, keep it.