spec(0.9.0): §3.5.2 said the reference producer does not emit param_histograms — it does

[coderoast-dev]

Change class: Editorial (GOVERNANCE.md §2) — editor merges, no version bump. No
normative sentence changes, no field's required/optional status moves, no document
changes verdict.

What was wrong

§3.5.2's Wire-emission status note said the reference producer (insight-metalog)
computes param_histograms, carries them through compose() in its internal
representation, and "does not yet emit them on the wire".

It emits them. The block is serialised inside every top_k entry that has tracked
slots, whenever the producer's per-template slot-tracking cap is above zero. That cap
is off by default — which is the true half buried inside the false one, and why a
consumer still has to apply §3.5's absent-array rule — but the shipped batch
configurations set it above zero, so published documents carry the block, and the
producer has a byte-exact wire test pinning its key-sorted shape.

This spec already disagreed with itself about it. §11's size budget prices
param_histograms as a live wire cost and names its two undeclared producer parameters
as residuals of the envelope. That paragraph cannot be written about a block nobody
emits. Two sections described one field's wire status in opposite terms, and a reader
sizing a document was reading the informative one.

The scheduling clause, and why it goes rather than moves

The note gated emission on a magnitude-aware ordinal metric landing first, on the
argument that js_divergence over value_counts reads a numeric slot's support as
unordered. The argument is sound; the gate is not what answered it. A slot the
producer declares ordinal is routed to a different carrier and never appears in this
block at all. That is now stated as an exclusion — a property of the block, which a
reader can check — instead of a schedule.

It should not have been a schedule in the first place. This is a vendor-neutral
standard: GOVERNANCE.md §3 makes the reference implementation an oracle, not a release
calendar the spec text waits on.

The half that stands

MetaLogDiff.field_histogram_deltas is computed on every diff and is not
serialised. It now sits in its own paragraph so it cannot be repaired away together with
the first half. It is a declared limitation, not a defect: §13.2 makes the member
optional, so a diff carrying any other signal field satisfies the clause without it, and
it lands with the full-fidelity (batch) diff surface for §11's reason.

What is preserved verbatim

• "When emitted, value_counts MUST serialise in a deterministic (key-sorted)
  order for replay bit-identity (§15.6)."
• The batch-vs-streaming reasoning (§11's envelope) — re-attached to the two halves it
  actually governs.
• The js_divergence order-blindness caveat, restated as the reason for the ordinal
  exclusion.

Checks

conformance/metalog_validate.py --selftest → SELFTEST PASSED: 16/16 fixtures, and
the shipped example validates CONFORMANT. Neither reaches prose; both are run because
the branch touches the repo the workflow gates.