Record why the embedding metadata copy is a repair net, not a fold

[bjmeetsfo]

Comment only — no behaviour change. It records a measurement so the trade behind
context_embedding rows is not re-derived from scratch.

What was investigated

Whether a context_embedding row's metadata could be folded onto its owner, so the rows
could eventually stop being written. There is nothing to fold: the metadata is already
on the owner.

Measured on a real ingest, comparing every row against its owner record:

fields IDENTICAL on the owner   29 distinct
fields unique to the row         4   (dim, model, model_ref, storage_options)

The four unique ones are encoder provenance. Everything else is a copy.

record_with_embedding_defaults() fills only fields the owner is missing, so retrieval
is already owner-first and this copy is never consulted in normal operation. It is a
self-repair net for an owner that has lost fields —
test_retrieve_recovers_hot_event_type_from_embedding_metadata builds exactly that case,
stripping event_type/classification/status/source_kind from a context_event and
asserting recovery.

The trade, measured end to end

dropping the copy   129,938 -> 126,510 bytes   (2.6%)
copy ON  -> repair test passes
copy OFF -> repair test fails

2.6%, because a row's bytes are the vector, not its metadata. Losing self-repair for that
is not worth it — and dropping the rows outright costs more still: eight tests covering
hot_event_type recovery, cross-session profile lineage and memory-layer classification
read them.

Why a comment and not a knob

A knob was implemented and measured first, then removed rather than shipped. An option
nobody should enable is the same declared-config-that-should-not-be-used shape already
found twice in this codebase — write_secondary_index, read by nothing until it was
wired, and raw_storage_policy, stored and displayed but branched on nowhere. The finding
belongs in the code, not in a switch.

Verified comment-only: every added line is a comment, no removals. Mirror imports clean.