🐛 fix(bash): escape the prompt for every PS1 state - #3466

Generate prompts from the tokens each shell reads as syntax, and run
each one through every prompt state and later action of that shell
with the shared matrix driver. Each example builds the activator in
process, so the search exercises quoting without a launch per example,
and per-shell seeds replay the known bypass shapes first.

Hypothesis rather than Atheris: coverage guidance adds nothing when the
system under test is a subprocess, and the property env runs on
Windows, where batch and PowerShell 5.1 live. The check job installs
tcsh, zsh, fish and nushell for that env on Linux; xonsh comes from the
property dependency group.

creator.py and pip_invoke.py opened the .venv redirect and the wheel
RECORD with a plain write, which follows a link planted after the
existence check. Both now go through write_file, which unlinks the old
entry and creates the new one with O_EXCL.

Windows refuses to unlink a read-only file, so write_file clears the
attribute with the clear_readonly helper safe_delete uses. write_file
now replaces a read-only redirect that points at a virtualenv on every
platform, where the plain write warned and left it.

Two queries matched code that no longer exists. UnquotedActivationValue
sank on str.replace inside instantiate_template, which the single-pass
substitution replaced, and MultilinePyvenvCfgValue sank on write_text,
which PyEnvCfg.write no longer calls. Both passed their fixtures while
reporting nothing on the tree.

Sink the substitution values on the iteration variable used outside a
quote() call, and sink write_file's data argument. Fixtures now mirror
the current shapes, with a positive and a negative class each.

Add four queries for code shapes from the October advisories: an
in-place write that follows a link, chmod inside an rmtree error
handler, a Python or xonsh quote that uses repr() instead of ascii(),
and a ScriptMaker that builds its shebang without shell_shebang().
Each ships with a fixture that has one positive and one negative case,
so the query test fails if the query stops matching that shape.

The in-place write query allows the app-data cache and the three wheel
image modules (base, copy, symlink). Those write into virtualenv's own
app-data cache, not into the environment being created, so this query
does not cover them.

packages.microsoft.com answered apt-get update with 403 and an unsigned
InRelease on one runner, which failed the job before any test ran. No
job installs from that repository, so the steps remove its sources
before updating.

Nushell and fish came from apt and brew, which hold a global lock, so they
waited behind the other packages in one step per platform. They now download
from their pinned GitHub release assets with a sha256 check, as RustPython and
the Windows Nushell build already did, and each runs as its own parallel step.
The fish PPA and the fury.io Nushell repository, two more apt sources that
could fail apt-get update, are gone.

setup-python finds GraalPy by listing the GitHub releases of
oracle/graalpython, and that REST listing now often returns an empty page
with status 200, so GraalPy jobs failed with "No release was found in GraalPy
version.json" or hung until the job timeout. uv installs every interpreter in
the matrix from the download list built into uv, as the 3.15 jobs already did,
so the tests no longer depend on that listing and the GraalPy pin goes away.

The artifact client retries HTTP errors but rethrows ETIMEDOUT, ECONNRESET
and the other network error codes on the first attempt, so one dropped
connection to the artifact service failed a job whose tests had passed.

The stdlib venv module run from inside a virtualenv on Python 3.10 and older
records the virtualenv's bin directory as home, so the new interpreter finds
its standard library only through the prefix compiled into the base build.
That prefix exists for the setup-python builds but not for the uv-managed
ones, which report /install, so the bare interpreter died with "No module
named 'encodings'". virtualenv points the environment at the base interpreter.

uv prefers an installed interpreter over a download, and the macOS runners
ship Apple's Python 3.9, so the 3.9 jobs tested that build instead of a
CPython 3.9 release. The Homebrew legs keep the default because they pass
Homebrew's interpreter through UV_PYTHON.

make_zipapp fetched each locked wheel with one urlopen call, so a single TLS
EOF from the index failed the zipapp check. The download now retries
connection errors, timeouts and 5xx responses twice with a short backoff; a
4xx response still fails at once, and the sha256 check is unchanged.