feat: the Weather COUNT runs while a staging maintains a copy

[krzysztof-smartdataengines]

Summary

The Weather starter's COUNT executor now runs while a staging maintains a copy. A staging keeps the
source authoritative, and the controller now carries saved queries through a staging, so the
approved COUNT stays current on the staging's map. The executor still refused any map with a copy
("Execute the count demo before staging or after completed cutover"), which left a current COUNT
that could be neither exported nor run until the cutover. The controller's installed acceptance of
queries surviving operations found this: its export-count-query after each staging was refused by
this rule, not because the COUNT was out of date.

What changed

• sde_demo.query_count._source names the source on a map that also maintains a copy. A query
  written for the copy is still refused ("does not describe this project's current source and
  map") until the cutover makes the copy the source.
• Relaxing the rule alone was not enough, and the new live test shows why: run_count_query opened
  its session with the source's adapter only, and sde.Session refuses a map that names an engine
  it has no adapter for (EngineError: the placement map refers to engines that were not supplied: ['clickhouse']). The executor now connects every engine the local map names, as the application
  does during a staging, counts on the source's adapter only, and closes every adapter even when one
  cannot close (ExitStack, one callback per adapter).
• docs/weather-starter.md: the rule, and the second connection during a staging.

Tests

• test_the_count_runs_on_the_source_of_a_real_staging[postgres|clickhouse] (live): setup, a run,
  a real staging through the starter's own operator stage with a signed packet, a second run
  writing through both copies, then the COUNT for the staging's map counts all 8 rows. The same test
  makes one adapter's close fail and checks that the failure surfaces and both adapters were
  closed. It was red before the fix with the EngineError above.
• test_a_count_reads_the_source_while_a_staging_maintains_a_copy[postgres|clickhouse] and
  test_a_query_written_for_the_maintained_copy_is_refused: the metadata path on the staging's map,
  loaded from a signed staging packet. They were red on the old rule with its message.
• make check locally on this commit with both live engines: ruff and mypy clean, Python 2015
  passed and 10 skipped (the orderbook slice, which needs the engine's C++ library), TypeScript 953
  passed.

Mutations on the committed tree, with both live engines: six killed - the old rule restored, the copy's
materialization accepted as the source, a session without the copy's engine, the count on another
adapter, adapters never closed, a cleanup that stops at the first failing close - and both controls
(two comments) survived.

The controller side, private, adds a test that export-count-query after a completed staging
exports the carried revision bound to the staging's map. It was red on main with the message the
installed acceptance recorded.

🤖 Generated with Claude Code