feat(emit): --help, a --panic step, and a gate on the step list

[vyncint]

Three issues about the fixture almost every integration test drives.

emit --help (#304)

The step language moves to fixtures/emit/src/steps.txt. The module header
includes it between its ```text fences and --help prints it
after a short banner, so there is one copy and the flag and the header
cannot say different things — which is what #304 asked for, and what makes
#318 possible at all.

$ emit --help
usage: emit [STEP]...

A termlens fixture: writes exactly the bytes its steps describe, in
order, then waits, sleeps, reads or exits as told. Steps apply left
to right.

TEXT             literal text — any argument that is not a step below
…
$ emit --nonesuch
emit: unknown step `--nonesuch`
emit --help lists every step          # exit 2, unchanged

--help is handled before any step is parsed, so it explains itself
whatever follows it, and it never becomes a Step.

--panic MESSAGE, and what it found (#311)

The README's "What TestBackend cannot see" table and skills/termlens/SKILL.md
§1 both promise a panicking child's message is assertable with
s.contains("panicked"). No fixture had ever panicked, so neither claim had
a test. #311 asked me to write down what I observe, not what I expect,
and to correct the README if the message did not survive.

The README is right. Measured three ways, all now asserted in
crates/termlens/tests/process.rs:

the child	what lands on the grid	status
panics plainly	the message, under what it had drawn	exit code 101
panics inside the alternate screen, no hook	the message, next to the frame it died on; alternate_screen() still true	exit code 101
tears the alternate screen down first (what a hook does)	the message on the restored primary screen; what the alt buffer held went with it	exit code 101

The status is an exit code of 101 — the Rust runtime's value — and not a
signal, which is the part the issue asked to take from observation rather
than assumption. No README change is needed.

The test uses a 100-column grid on purpose: the runtime's own
panicked at <file>:<line> line wraps on a narrow one, and a wrapped needle
would make the test about the width instead of the panic.

check-emit-steps.sh (#318)

steps.txt and --help are one file, so they cannot drift. What they could
still drift from is the match that parses the steps. The script compares
the documented set against the implemented arms and names the difference in
either direction. It runs in the gates-listed job — no toolchain, just two
greps and a comm — and is listed in CONTRIBUTING §1 with the other scripts.

Proven to fail before being trusted to pass:

$ # a documented step with no arm
::error::`--ghost` is documented in …/steps.txt but no `"--ghost" =>` arm implements it
$ # an arm with no documentation
::error::`--ghost` is implemented in …/main.rs but not documented in …/steps.txt
$ # a rename on one side — both halves reported
::error::`--windowsize` is documented in …/steps.txt but no `"--windowsize" =>` arm implements it
::error::`--winsize` is implemented in …/main.rs but not documented in …/steps.txt

It also refuses to pass when either pattern matches nothing, so a future
edit that breaks the extraction cannot quietly turn the gate into a
tautology — the failure mode a grep-based gate actually has.

On main today: 28 documented, 28 implemented, the same set.

bash 3.2 clean: no declare -A, no GNU-only flags, and the comm output
goes through files rather than a process substitution so the loop's
status=1 is the one that gets read.

Verification

• cargo test --workspace --all-features green, fmt, clippy … -D warnings,
  RUSTDOCFLAGS='-D warnings' cargo doc --no-deps -p emit clean — the last
  one matters because the module header is now an #![doc = include_str!].
• check-ci-gates-listed.sh green with the new line in CONTRIBUTING §1.
• emit --help, an unknown step, and --panic all exercised by hand at a
  prompt before any test was written; the panic behaviour was observed
  through termlens inspect first.

What this does not do

• No CHANGELOG entry. Nothing here ships: fixtures/emit is
  publish = false and the script is repo infrastructure.
• zizmor was not run locally — no pipx on this machine. The ci.yml
  change is one - run: with a literal path, no expressions and no new
  permissions, so the zizmor job is the check on it.
• The thirteen other test files that say "steps are documented in
  fixtures/emit/src/main.rs" are left alone; they still resolve, since the
  header is in main.rs. Only tests/common/mod.rs, the shared helper the
  rest echo, was repointed at steps.txt and the new gate.
• --help is not itself in steps.txt and not counted by the gate: it is not
  a step, and its arm does not have the shape the script matches. The script
  says so where a reader would otherwise wonder.

Closes #304
Closes #311
Closes #318