Raise the core tmux minimum to 3.3a - #21

why: tmux 3.2a sized detached windows from the command client unless
`window-size` was manual, so the create and guide fixtures forced that
option on that release only. The 3.3a floor removes the release.

what:
- Remove the forced option from the create and guide fixtures and the
  note from the create guide.
- Keep reading the daemon version in the guide fixture for its deferral
  check and in Create.Tests.ps1 for tmux 3.4's dollar-sign escaping.

why: The workflow still exercised 3.2a, which the module no longer
supports, and had no lane proving that older daemons are refused.

what:
- Pair PowerShell 7.4.20 and 7.6.6 with tmux 3.3a and 3.7c, and swap
  3.2a for the same 3.3a source commit in the advisory macOS trial. The
  MCP discovery step follows the lower endpoint.
- Add a `reject` job that builds tmux 3.2a and unpatched 3.3 from their
  pinned source commits, checks the tag, and runs only the VersionCheck
  suite against the packed modules. The 3.3 commit is the `3.3` tag of
  tmux/tmux; the shared build script does not know that release, so the
  job builds from source directly.
- Add the job to the aggregate gate and the release artifact binding, so
  a release needs the rejection cells as well as the supported ones. The
  artifact checker's job set and its unit test follow.
- Add a `reject-old-daemon` job that builds tmux 3.3a and a 3.0a or
  3.1c daemon from their pinned tag commits and runs the VersionCheck
  suite with the 3.3a client against that daemon. The job joins the
  gate and the release artifact binding.

why: The install note, the compatibility page and the contributing guide
disagreed about the minimum tmux.

what:
- State the 3.3a minimum, how the connected daemon is checked, and where
  it is tested in the README, the compatibility page and the
  contributing guide.
- Explain on the compatibility page that the linked .NET package checks
  the daemon before it sends anything, that operations fail with
  `TmuxVersionTooLowException` instead of returning nothing, what the
  command guard does when a daemon is replaced, and that the control
  cmdlets need no check of their own.
- Prepare a 3.3a plan in the Get-TmuxQueryPlan example; regenerate the
  help XML and follow in the help-example and source query tests.
- Describe how daemons older than 3.2 are refused and which CI job
  proves it.

why: Several suites wrap tmux in a script that logs one line per
invocation, then assert that invalid input, previews and offline paths
dispatch nothing. The version check now runs one hook-free probe, or an
executable `-V` for an absent endpoint, ahead of the library call that
rejects the input.

what:
- Make those wrappers skip exactly the probe and `-V` forms and keep
  counting every other invocation, including the fenced batch.

why: tmux 3.3a's configure stops with "must give --enable-utf8proc or
--disable-utf8proc" when it finds utf8proc, which Homebrew installs, so
the shared build script's bare configure failed every 3.3a cell before
any suite ran.

what:
- Configure the already cloned and verified source in the trial with
  `--enable-utf8proc`.
- Write the same source-commit marker the later steps read, and keep the
  3.7c cells on the same path.

why: With `--enable-utf8proc` the arm64 cells stopped at "utf8proc not
found": Homebrew installs under /opt/homebrew, outside the compiler's
default search paths, while the Intel cells find it under /usr/local.

what:
- Pass the include and library directories from `brew --prefix utf8proc`
  to configure.

why: Readers could not tell how a refusal behaves under `-ErrorAction`,
whether the MCP server is covered, or why the benchmark receipts name
3.2a.

what:
- Say on the compatibility page that a refusal is a non-terminating
  error that follows `-ErrorAction`, so `SilentlyContinue` hides it as
  it does any operation failure.
- Say that `kill-server` on an absent endpoint is a no-op, and that the
  separate MCP tool is built on the same package and refuses the same
  daemons.
- Note in the benchmark README that its 3.2a receipts predate the floor.

why: The Product suite stops at its first failing worker, and on macOS
that is ResourceInstall.Tests.ps1, so the version check suite never ran
there.

what:
- Run `-Suite VersionCheck` first in the advisory trial and report it
  beside Product and Documentation in the summary and warning.
- Keep `continue-on-error` on every step: the trial stays advisory.

why: The compatibility page and the contributing guide listed an
outdated matrix.

what:
- List the releases the Linux workflow now exercises.

why: On tmux 3.7 and newer alpha.22 sends one-shot reads over a shared
control client. Scripts that wrap tmux to trace or inject invocations
saw none of them and failed with the new pin. The other scripts do not
depend on seeing each command and run on the default transport.

what:
- Set LIBTMUX_CONTROL_TRANSPORT=0 in eng/Test.ps1 only for the Snapshot,
  Attachment, SourceQuery, WorkspaceApply, WorkspaceCorpus, Help and
  GuideExamples scripts, which wrap tmux.
- Say so in CONTRIBUTING.

why: alpha.22 attaches no control client for a pane text wait below tmux
3.7. The wait polls without -AllowPollingFallback and reports
PollingFallback, where alpha.20 refused to poll without the opt-in. The
checks that expected no polling failed on tmux 3.3a through 3.6.

what:
- Add Test-PollingExpected, which reads the tmux banner, and expect
  PollingFallback below 3.7 in the pane text wait, README, guide,
  help-example and MCP discovery checks.
- Say in the input guide, the Wait-TmuxPaneText help and the MCP guide
  that older releases always poll, and regenerate the help XML.
- Say in the MCP guide that waits require control observation on tmux
  3.7 and newer.

why: alpha.22 checks the daemon before it sends a command, so a stand-in
transport meets a version probe and a guarded command list, and the
wait-context probe failed with "the wait channel could not be
withdrawn". The package also verifies exact source evaluation only for
stable releases up to 3.8, so a later patch release or a development
build refuses the exact plan the source query suite required.

what:
- Answer the probe and unwrap the command guard in the wait-context
  probe's stand-in transport.
- In the source query suite, expect the planner's refusal and
  residual-only plans on a tmux that is not source-verified, and run
  every other check there.

why: A cmdlet that surfaced the package's TmuxVersionTooLowException
reported it as an invalid operation, like any failed command, so a
script filtering on the error category could not tell an unsupported
tmux from a failure. The package's remedy names LibTmux 0.0.0-alpha.21,
which is not a module a PowerShell user can install.

what:
- Report TmuxVersionTooLowException as ErrorCategory.NotImplemented.
- Replace the package's remedy in the error details with the release a
  PowerShell user can install, LibTmux 0.1.0-alpha2 from the PowerShell
  Gallery, and append it if the package text changes.
- Look through inner exceptions for the refusal, so one wrapped by a
  workspace build failure gets the same category and details.

why: The package checks the daemon's version before it sends a command.
Nothing here fails when a cmdlet stops reaching that check, a refused
command takes effect, or the check starts firing tmux hooks.

what:
- Add tests/VersionCheck.Tests.ps1, which runs against the tmux first on
  PATH.
- Below 3.3a, assert that session, client, hook and snapshot reads,
  Connect-TmuxControl, New-TmuxSession, Set-TmuxOption,
  Set-TmuxEnvironment, Wait-TmuxChannel, raw commands and chains are
  refused on a live daemon and on an absent endpoint, including through
  a client wrapper that claims `tmux 9.9`, and that the daemon's
  sessions, option, environment and the absent socket are unchanged.
  Every operation must be refused as NotImplemented with a remedy that
  names LibTmux 0.1.0-alpha2. Discovery operations on an absent
  endpoint, which start nothing, must report the missing socket in a
  TmuxCommandException; a read on a supported executable reports the
  same.
- Capture a pane and a workspace plan through a double that claims a
  supported tmux, then replace the daemon with one below the floor and
  assert that Wait-TmuxPaneText and Invoke-TmuxWorkspace are refused as
  NotImplemented with the remedy and that the host script never runs.
- Check the remedy rewrite directly, for the package's wording and for
  a reworded message that gets the release appended.
- On 3.3a or newer, assert that reads and writes are checked and that
  neither the check nor the fence fires an after-display-message or
  command-error hook, measured on a handle not yet checked so the probe
  itself runs, with a native display-message and a failing command as
  positive controls. The command-error half is reported
  SKIPPED before tmux 3.5, where `set-hook` rejects it.
- Assert that a replaced daemon fails the fence for both batch cmdlets
  and leaves a sentinel session uncreated, that a replacement below the
  floor is refused, that a supported replacement is checked again
  before the command runs, and that an unchanged daemon behind the same
  wrapper is not refused.
- Assert the version grammar: 3.2a, 3.3, next-3.3, 3.3-rc, master and
  malformed text are refused; 3.3a, 3.4, 3.4-rc, next-3.4 and 3.7d are
  accepted.
- Assert that one server handle probes once, that a daemon gone after
  the check is checked again before the command is sent again, and that
  a renamed executable still starts a daemon.
- On tmux 3.7 or newer, assert that attaching and commanding through
  the control cmdlets fires no more hooks than the package's own attach.
- Answer the probe with doubles that print the 3.1c usage line, the 3.0a
  protocol refusal and the unknown `-N` error, and require each refusal
  to name its release and the remedy. The 3.1c and 3.0a doubles keep
  stdout open past the client's exit, as those daemons do, and the read
  must fail within 10 s. With LIBTMUX_OLD_DAEMON set, the same bound
  applies to a real 3.0a or 3.1c daemon behind a 3.3a client.