feat(semantic): outbound chokepoint (yesod) + retention probe with an exercised rollback

[mdheller]

Closes the two remaining kernel gates SEMANTIC_CONTROL_ARCHITECTURE.md listed as not
built. Rebased onto current main, so it sits on top of #128–#130.

serialization_channel — the yesod chokepoint

emit is the only function that produces a WireEnvelope, and it refuses unless
the caller presents an InternalModelState whose equilibrium clears.

That is the join between the two organs: daat decides, yesod transmits, and
transmission is not reachable without a decision. An agent cannot route around its own
share/withhold equilibrium by serialising somewhere else, because there is nowhere else.

This deliberately inverts the usual arrangement, where a policy check is something a
call site remembers to perform. Here the check sits on the only path out, so forgetting
it is not expressible. The refusal raises rather than returning falsy — a refusal a
caller can ignore by not checking the result is not a refusal.

The invariant is checked against the real tree, not asserted in prose:
single_channel_violations fails the build on JSON encoded outside the encoding
primitive, or a WireEnvelope constructed anywhere but here. Verified against all 16
kernel modules including the VSA/VRF/interferometry/twin work from #128–#130 — still
one path.

An address travels as a skeleton(), never in full: structure crosses, the referent
and evidence pointer stay behind.

retention_probe — anti-forgetting, with a rollback that actually runs

incorporate probes retained tasks → applies → probes again → returns the original
object if any retained task regressed beyond tolerance.

A rollback that is "return what you already had" cannot be incorrect, which deletes the
category of undo bugs rather than testing around it. The architecture record was
explicit that until the rollback runs in CI the guard would be decorative — the
rollback path is now the most heavily exercised thing in the file.

Two refusals worth calling out:

• an empty probe refuses rather than clearing. A gate that passes everything is
  worse than no gate, because it looks like one.
• probes disagreeing on the task set raise. A task that vanished between probes is
  a broken probe; reporting it as "no regression" would turn the most alarming possible
  result into the safest-looking one.

A real bug the threshold test caught

1.0 - 0.02 is 0.98, whose loss fraction is 0.020000000000000018. A bare >
therefore reverted at exactly the documented tolerance. The isclose guard is
load-bearing, not decoration. Fixed in the implementation rather than by softening the
test.

Testing

67 new tests, every guard on its refusal path. 583 green overall.

Still not built

Fibration depth budget, and hod/malchut wiring. The depth budget deliberately waits
on fibration delegation actually existing — enforcing a budget for a mechanism that
isn't there would be a control with nothing behind it.