[JimmyWesley]

Release 0.65.0 — the retrieval is done long before the reply is

A hosted answer is two costs three orders of magnitude apart. The sweep's
retrieval is milliseconds — 19 on the 82-node fixture, 448 on a 1,902-node
forest cold — and the provider round trip is seconds. J.10.11 already keeps
the model off a reader lane and J.10.6 already publishes the two halves apart,
but the Station still answers once: the bundle exists as a value in the host
at millisecond 19 and leaves the building at second 10.

The Ask console spent that gap on a spinner.

What changed

The path panel (J.5.15). The console runs the sweep's retrieval itself, in
parallel with the answer, through the ordinary harvest — same question, same
k, same entry ranker, read-only and deterministic — so what it draws is what
the answer will see rather than a guess at it. It deposits no pheromone: heat is
the whisper's at the close of an answer (J.10.7), never a read's. The map is the
J.11 graph projection Explore already reads, already scoped and already
filtered.

Two of its rules are where the honest version and the flattering version part:

	what a demo would draw	what it draws
sweep	a cited stage, "the model chose these"	none — evidence is the whole bundle and the reply is prose that names nothing, so there is no selection to show
walk, entry stage	filled from sources, to look complete	zero — J.10.4 keeps only what carries text, and zero is what actually happened

The switch is deliberately separate from hops: a walk costs one model call
per hop and the drawing costs none, so one control over both would teach an
operator that the picture is what made the answer slow.

The panel prints the real millisecond figure off the Part D trace beside a
reveal that takes seconds. A console whose subject is that retrieval is cheap
must not leave an audience reading the animation's duration as the measurement.

A boundary under the shell. React 18 unmounts the whole tree on an uncaught
render throw, so one bad element below App left a blank page — no navigation,
no forest picker, no way back. Not hypothetical: a <Field> wrapping a
<Select> made React refuse an <input> with children (#137) and took the
Studio down every time a document was opened. That bug is fixed where it was
written; the boundary is the answer to the class. It sits below the Shell and
never above it — a boundary that also caught the navigation would leave nothing
to navigate with.

The reading console is a search. Each result is a real href, so middle
click and "copy link" work. Three honesty rules: the clock is the engine's
(Server-Timing: vine, with the round trip in the tooltip and no time at all
when the header is absent); the page is one page (locate has an 800-token
budget and no cursor, so the line says there are more matches instead of letting
ten look like all of them); and the search names the search it did not do
(locate reads curated metadata, never bodies — a silence here is a fact about
titles and summaries).

Not here

The walk's hops arriving one at a time, live. That needs the host to push, which
is a contract, and this version does not add one.

Engine

src/monkeyllm/ is untouched — the wheel's module code is identical to
0.64.0's. The number moves because it is the Station's identity, not the
engine's: package_version() feeds serverInfo.version (J.1.2 r3), the
station: field of forests() (r6, which the generated skill stamps and
teaches operators to compare), and /v1/health. A host change that does not
bump it leaves every deployment reporting a build it is no longer running.

Verification

• Full suite green, both legs.
• python -m build + twine check pass on the sdist and the wheel; the
  workflow's tag/version guard simulated OK against v0.65.0.
• check-trail.mjs is F.137 and runs from tests/test_trail_panel.py, the same
  arrangement check-skill.mjs has.
• de599e4 repairs a self-inflicted one: trailmap.js was first committed with
  NUL bytes, which made git treat it as a binary blob and cost the file its text
  diffs. tests/test_repo_hygiene.py now guards the class.

Known gap

The v0.65 changelog block in the spec carries the prose but not the closing
bullet list nor the Acceptance: F.137 line that every version since v0.54
ends with. F.137 itself is present in the body of the document.