feat(stream): read the connection ack instead of assuming it

[htcom-code]

The platform opens every stream connection with a ready frame now
(protean #80, 237bcd0), stating what it is, what it is
recording, how often it pushes and how much it retains. The console had
been asserting four of those on its own. It reads them instead — and
where the platform says nothing, it says it does not know rather than
falling back to the guess.

Two further fixes came out of staring at an idle platform while
verifying this, and are in their own commits.

What the console stopped asserting

	before	after
LIVE badge	LIVE · 200 buffer on every platform — a ring size this console had never been told and could not see	the stated capacity, or no claim at all
silence watchdog	6s, because the platform was assumed to push at 1Hz	6 × the stated tickMs, clamped to 2–60s
module table, empty	"enable protean.trace.metrics.enabled=true and restart" — told operators to switch on a setting that was already on, and the key is live, so "restart" was wrong too	reads metricsEnabled: off / on-and-idle / not stated
trace table, empty	"requests appear here as the platform serves them" — a promise nothing keeps while trace.enabled=false	reads tracesEnabled and says recording is off
header	origin only	origin · protean 0.1.0-SNAPSHOT (display only)

tracesEnabled is read before metricsEnabled in the module panel,
and that ordering is required rather than defensive. Per-module metrics
are aggregated from recorded requests, and the contract states
metricsEnabled as the effective value (trace.enabled AND
trace.metrics.enabled) — so a platform with recording off reports
metrics off as well, and that value read on its own produces "turn it on
and rows appear": an instruction that changes nothing beside a promise
nothing can keep.

The ack is display and diagnostics only — nothing branches on which
implementation answered. Its values are connection-scoped and dropped
when the stream is rebuilt, because they are live settings on the
platform and the last connection's answers may no longer be true.

A platform that sends no ack is unchanged. Go still sends the older
hello, Rust sends nothing, and both captures stay in the fixtures as
that axis: the console stays live, claims nothing, and keeps the 6s
window it has always used.

Verified against a running platform, not only against tests

examples/quickstart on :8080, protean 237bcd0. The protean fixture in
stream-fixtures.ts is the real connect bytes from that session
(correlation ids synthetic — public repo), not a constructed example.

• ready first, once per connection; metrics/modules/summary
  then repeat at 1Hz. buffered (5) equalled the rows of the trace
  frame behind it (5).
• The three empty-state screens were each produced by restarting the
  platform into that state, because they cannot be told apart from the
  data: the metrics frame is [] in all of them.
  • metricsEnabled: true → No module traffic yet
  • metricsEnabled: false → Module metrics are off
  • tracesEnabled: false → Trace recording is off, in both panels
• Reconnect re-learning was exercised for real: a platform restart
  dropped the stream, the hook rebuilt it, and the panels changed on the
  new ack without the page being touched.

The contract dependency above was found during that verification and
fixed on the platform side as well: the ack was sending the raw metrics
switch while the MCP tool answered with the effective value, so the same
name meant two things. Reported, and protean has pinned the contract to
the effective value.

Also in this PR

fix(ui): keep an idle screen readable and truthful — the latency
chart removed itself when there were no samples, taking half the row it
shares with the status mix and leaving a layout that read as broken
rather than quiet. It draws its frame now, on a scale labelled as a
default, with nothing over the grid: a line, an area or a point would
each be data the platform never sent. And the status mix printed
1 req beside four zeroes — the header count and the bar widths were one
number, and the || 1 that keeps a width from dividing by zero was also
what the header showed.

fix(dev): resolve the proxy target once, for both uses — the dev
proxy read process.env, which does not carry .env files (Vite loads
those itself), so setting VITE_PROTEAN_TARGET in .env as the README
instructs moved the name in the header while every request still went to
the default: the header named a host nothing was sent to. Unset, the two
then fell back differently (proxy to localhost:8080, header to the dev
server's own origin). The config resolves it once and hands it to both.
Dev only — a production build carries no proxy and must keep naming its
own origin, which is where its requests actually go.

Tests

199 passing, 29 of them new: the ack's watchdog sizing and clamps,
partial and unreadable acks, per-connection scoping, the three empty
states in both panels, the empty chart, and the status count. The
ack-contradicting pair (tracesEnabled: false with metricsEnabled: true) is kept as a guard — the console talks to three implementations
and cannot verify an ack, so it declines to promise rows on the strength
of one.