fix: report an exec failure as a spawn error, not a SIGABRT child

[vyncint]

What & why

On Unix, a program that exists but cannot be executed made spawn return
Ok with a child already dead of SIGABRT. termlens inspect drew a screen
holding fatal runtime error: assertion failed: output.write(&bytes).is_ok(),
reported exited: killed by signal: Aborted, and exited 0. portable-pty's
pre_exec closes every descriptor above stdio, and std's close-on-exec
exec-error pipe is one of them (upstream: wezterm/wezterm#7893, still open).

On Unix termlens now starts the child itself (src/spawn.rs), building the
command portable-pty 0.9.0 would build: the same resolution, argv[0],
working directory (with portable-pty's fallback to the child's HOME),
environment, session and signals. Descriptors above stdio are marked
close-on-exec instead of closed, so a failed exec reports through std. A
unit test starts one probe through both spawns and requires the same
answers (argv[0], directory, environment names plus a checksum of the
values, descriptors, umask, process group, controlling terminal), so a
portable-pty update cannot drift away unseen.

The error names the culprit where the errno would point at the wrong file:

cannot execute /w/./badinterp: its #! interpreter "/nonexistent/interpreter" does not exist
cannot execute /w/./crlf: its #! interpreter "/bin/sh\r" does not exist
cannot execute /w/./x: its #! interpreter "/w/plain-text" is not an executable file
cannot execute /w/./noloader: No such file or directory (os error 2); the file exists, so what is missing is its dynamic loader
cannot enter /w/locked: Permission denied (os error 13)

A failure before the exec (a NUL in a variable, the fork) reads as std
reports it, as before. Four rare setups still go to portable-pty unchanged,
and docs/DESIGN.md lists them: children reaped by the kernel (std would
panic), no usable directory and no HOME, a program that does not resolve
(portable-pty's own diagnosis), and a slave that cannot be opened by name.

Also in this PR:

• dup_writer duplicates the master with std's close-on-exec
  BorrowedFd::try_clone_to_owned. Its plain dup leaked into every other
  child the test process started while a terminal was alive.
• XTGETTCAP's TN reports the last TERM set, the one the child gets.
• SECURITY.md lists the five unsafe blocks: borrow_raw, access(2),
  a sigaction(2) query, kill(2), and the pre_exec hook. DESIGN §1
  gains "On Unix termlens starts the child itself", which covers the
  lifecycle lock now held through the exec and the openpty close-on-exec
  window in the PTY layer. AGENTS.md and CONTRIBUTING §3 point at
  spawn.rs, and the skill's failure table gains the new message.

Closes #563

Review

The code-review skill ran at max effort on the first version and returned
15 findings. All of them are addressed in 95f1e77, or documented where they
are inherent:

• The lock is held through the exec, about 0.25 ms per spawn on Linux.
  That is the price of reporting the exec at all.
• The openpty window belongs to the PTY layer.
• The PATH snapshot only matters under an unsound concurrent set_var.

CI then found two problems in the new comparison test itself, both fixed:

• a read could stop between the marker and its line ending (edef00d);
• on macOS the test waited on the shell without draining the terminal, and
  a process exiting there waits for its queued output (the echoed answer)
  to be read, so every macOS stress shard hung (088aca6, b6c101a).

Each probe now closes its PTY under the lifecycle lock and runs under a
deadline. A hang reports ps and stty state instead of stalling the job.

Tests

Each was checked to fail without its part of the fix:

• process.rs: a missing interpreter, a CRLF interpreter, an interpreter
  that is not a program, an execute-only script, an E2BIG argument, a NUL
  in a variable and an unenterable directory.
• descriptors.rs, a binary of its own: no termlens descriptor reaches an
  ordinary child, and none of the test's reaches a termlens child.
• inspect.rs: the example and termlens inspect both exit 2 on a missing
  interpreter.
• queries.rs: TERM set twice.
• spawn.rs unit test: the two spawns cannot be told apart. It fails on a
  changed argv[0], a missing setsid, or a changed variable value.

Verification

• Behaviour diff: published 0.11.5 against this branch on 969 cases,
  covering render in five formats over 104 saved screens, diff pairs, the
  CLI surface, and 35 inspect programs. 967 are identical; the two that
  differ are the missing-interpreter and CRLF cases. Also checked by hand:
  the SIGCHLD-ignored case matches 0.11.5, and a deleted working directory
  runs in HOME as before.
• Gates: clippy -D warnings for Linux, x86_64-pc-windows-msvc and
  aarch64-apple-darwin; the MSRV checks at 1.85.1; the full suite, 625
  passed and 0 failed; preflight green on the previous head.

Checklist

• Linked an issue (or explained above why none exists)
• Tests added/updated for the change
• cargo fmt --all and cargo clippy --workspace --all-targets --all-features are clean
• All commits are signed off (git commit -s) — see CONTRIBUTING.md §5
• No AI attribution trailers (no AI co-authors, "Generated with" footers, or bot identities) — see CONTRIBUTING.md §6
• CHANGELOG.md updated under [Unreleased] (user-facing changes only)
• Snapshot changes (if any) were reviewed with cargo insta review, not blind-accepted