[taras]

Repairs #862. Main is red at f5f4f2d8 (run
36786555733),
and this is the whole of the failure: one test row, no product job.

Not Closes #862: that issue says it closes itself when CI next succeeds on
main, and closing it at merge — before the repaired head has been proven —
is exactly what it asks nobody to do.

Why

The Slice D conversation-order row waits for the two turns that report a
provider conversation, then expects clearing the filter to restore three turn
rows. A turn scripted queued is held before it produces anything: it never
reports a conversation, so it never contributes to that wait — but its row
exists as soon as its Prompt is observed. The row waited for one fact and
depended on the other.

What changes

Before:

• test-deno (1/15) failed on main, reading two Prompt slots
  (build · streaming, review · completed, recorded) where the row expected
  three. green failed with it.

After:

• The row waits for all three Prompts to be observed before it counts them, so
  what it proves no longer depends on which <Spawn> the scheduler reached
  first.

     yield* until(
       session,
       "both started turns reporting their conversation",
       () => started(session) === 2,
     );
+    yield* until(session, "all three Prompts being observed", () => observed(session) === 3);

The existing wait stays: the ordering assertions need those two conversations.
The new wait is independent, and says the queued Prompt's slot exists.

What must stay true

• The conversation-order assertions still concern only the two turns that
  actually reported conversations — this adds a precondition, it does not widen
  what the row accepts.
• No production file changes. The diff is six lines in one test, five of them a
  comment.

How to verify it

deno task test packages/cli/tests/repl-agent-interface.test.ts   # 9 passed, 31 steps, 0 failed
deno task check                                                  # 0
deno task lint                                                   # 0
git diff --check                                                 # 0

The failing schedule is not reproducible on demand locally — every local run
observed all three spawns first, which is why the row passed through four
review rounds. What this change removes is the dependency on that order, rather
than a defect that can be made to appear at will.

I audited every other toHaveLength(3) in the file: the rest are already gated,
either by observed(session) === 3 or by waiting on all three coroutines
finishing. This row was the only one.

Scope

Included

• One readiness wait in U1: conversations are ordered by their earliest turn, and selecting one changes only the filter.

Intentionally unchanged

• Every other row, every fixture, and all REPL production code.
• test-weights.json, dependencies and deno.lock.

Risks and limitations

• This restores main health; it does not measure weights. test-weights.json
  is three slices stale and two REPL suites are unmeasured, which is separate
  work.

Scope confirmation

• Every changed file supports the purpose described above.
• Unrelated cleanup and formatting changes are excluded.
• The description matches the final diff and test results.