test: replace the shell with the emit fixture at 72 of 99 sites

[vyncint]

Closes #249. Step 1 of #149.

What this is

fixtures/emit — a std-only program that writes exactly the bytes its arguments describe, then waits, sleeps or exits as told. One argument per step; the steps are the crate doc of fixtures/emit/src/main.rs:

TEXT  NL  CR  --text WORD        literal bytes
--esc BYTES  --csi BYTES         ESC …  /  ESC [ …
--raw SPEC                       \e \n \r \t \a \\ \xNN
--sleep DUR  --wait  --exit N    the only timing, the instant-exit guard, the status
--echo-line  --echo              read a line and write it back / copy stdin until EOF
--seq N  --cwd  --pid            the flood, the directory, the process id
--env NAME  --environ            one variable or all of them, sorted
--loop                           run what precedes once, what follows forever

tests/common::spawn_emit(builder, steps) runs it from any builder. Then the migration, one test file per commit so each diff is one file's scripts against their emit form: tabs, styles, charset, utf8, state, frames, process, builder_validation, observe, timeouts, stable, basic, concurrency, scrollback, input, queries, fixtures. Where a test used to be one printf, it is one --raw, with \033 read as \e — a reviewer diffs the escape sequences directly.

The count

grep -c 'spawn("/bin/sh")\|spawn("sh")' crates/termlens/tests/*.rs: 99 → 27, and every one of the 27 has a comment or file header saying why it is a shell:

file	left	why
queries.rs	13	read a reply byte for byte under stty -icanon -echo
input.rs	7	read the wire (mouse, paste, cursor keys) under stty
drain.rs	3	flood + raw mode; the deadlock the harness must never cause
backpressure.rs	2	dd an exact byte count of replies under stty
basic.rs	1	kill -TERM $$ — a process signalling itself
process.rs	1	trap … TERM — a signal handler

The 25 raw-mode sites share one blocker: reading what termlens typed back needs -icanon -echo, and std has no tcsetattr. That is the fixture's stated rule (std only, no timing but --sleep) doing exactly what it was written to do — the shell is the honest tool for a test that is about the line discipline. If those 25 should go too, the step is --raw-mode on the fixture via libc (one tcsetattr) or crossterm, and that is a decision about the rule, not about the migration. Flagging rather than deciding.

Two tests changed shape on purpose

• basic.rs send_str_and_enter_round_trip… waited for got: hello, which the line discipline's echo of the typed hello could satisfy on its own. It now waits for hello back — the fixture's read plus a suffix the echo cannot produce.
• builder_validation.rs a_bare_program_name_under_env_clear… used sh as the bare name and /usr/bin:/bin as the PATH remedy. It now uses the fixture as the bare name and the fixture's own directory as the remedy: the same refusal, no Unix path baked in.

Everything else asserts what it asserted before.

Verified

Every gate in CONTRIBUTING §1 locally, including the cross-target clippy; full suite three feature configurations. The windows and stress workflows are dispatched against this branch — results in a comment below, since #249's done-when names the stress workflow and #149's step 4 is what the Windows count is for.

[vyncint]

windows.yml against this branch — run 34179522277

345 passed, 92 failed — against 274 / 163 on main after #275. Of the 92, 30 name a shell (the 27 sites kept on purpose, plus the two wait_frame doctests and the tiocgwinsz python one); the other 60 are tests that used to die on /bin/sh before they could say anything and now run far enough to hit ConPTY. They line up with the step-3 table on #149 — every frames.rs, graphics.rs, and queries.rs reply test; mouse modes, focus, OSC 8 ids, DECSTR, DEC Special Graphics in charset.rs, the raw bytes in utf8.rs — with one new finding worth adding to that table:

• Child exit is not EOF on ConPTY. a_departed_child_is_refused_identically_on_every_platform, typed_input_to_a_departed_child_is_a_typed_error, a_mouse_click_at_a_departed_child_blames_the_child, resize_after_the_child_exits_is_refused_like_send and waits_fail_fast_on_eof… all fail the same way: the child has exited, and the harness has not seen EOF, because the output pipe stays open until the pseudoconsole is closed. On Unix the liveness check reads EOF off the master; on Windows it would have to read the process handle. That is a harness change for step 4, not a test change.

Nothing here is a regression on Unix — the suite, all three feature configurations, is green on the PR — and nothing here is the fixture misbehaving on Windows either: emit built and ran there, --wait blocked on cooked-mode input as read did, --environ reported the environment, and the two bin!-style absolute-path spawns work. The stress run is dispatched too; numbers when it lands.

[vyncint]

stress.yml against this branch — run 34179524611

All ten shards green: ubuntu and macOS at 1, 2, 4, 8 and 16 threads, the suite repeated per shard as the workflow weights them. That is #249's last done-when — the stress workflow passes on the migrated suite — and it is the one that mattered most, since the shell's process model was what leaked into the wait-semantics tests before (#243).