The streaming render tests drove real Rich, so the Live construction kwargs
and per-update refresh/flush flags weren't asserted and survived mutation.
Inject a fake Live to pin screen/auto_refresh/transient/redirect_* and the
forced refresh, and a flush-recording stream to pin status-notice flushing.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

- output: pin data_table pad_edge / detail_table padding, and assert emit_ndjson
  writes one flushed newline-terminated record
- context: assert resolve_session raises when only one of session/account_id is
  present (pins the `or` guard) and that a non-rejection NotAuthenticated still
  auto-logs-in with an env key set (pins the `and` in _should_auto_login)
- transcribe_render: assert exact sentiment percentages, mm:ss formatting across
  a 60s boundary, and most-relevant-first ordering for topics/content-safety

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

The transcribe --show-code tests parsed/ran the output but never asserted the
4-space indent of the rendered config kwargs, so the indent literal survived
mutation. Assert the exact indented config block.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

_survives wrote each mutated module via ast.unparse and ran the covering tests
in a subprocess. Consecutive mutants unparse to files differing by a single
token, so they're typically the same byte length and can be written within one
mtime-second. CPython validates a cached .pyc by exact (mtime, size) match, so
the subprocess could load the previous mutant's (or the original's) bytecode and
execute unmutated code — reporting a false "survivor" and failing the gate
spuriously (and flakily). Drop the module's cached .pyc after writing the source
so the subprocess always recompiles the mutant under test.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

- sources: pin _is_streamable_wav's full mono/16-bit/16k `and` chain, the
  file-not-found/ffmpeg-missing/empty-audio exit codes, that ffmpeg isn't
  terminated on a clean EOF, and the exit-code fallback when stderr is empty
- microphone: pin RawInputStream channels/dtype, the ~100ms blocksize floor,
  the `rate > 0` boundary, and that resample treats audio as 16-bit mono PCM
- client: pin the single-row validation probe limit, the verbatim-vs-fallback
  transcribe error message, and that a provided on_begin is wired to Begin

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

- config_builder: pin _derive_kind's Optional unwrap + bare-scalar/dict origin
  classification, and that KEY=VALUE / NAME:VALUE split only on the first
  separator (values may contain '=' / ':')
- youtube: assert yt-dlp is driven quietly and actually downloads, and pin the
  no-file-produced exit code
- ams: assert the error "detail" field is extracted cleanly rather than leaking
  the raw JSON body (the fallback happened to contain the same substring)

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

- agent/audio: pin the device-rate blocksize, exact pending() sample count, the
  callback's exact zero-fill remainder, the audio-open exit code, and that
  close() lets the stream reopen on a later start()
- agent/session: assert the server error message wins over code/fallback, that a
  transcript without "interrupted" defaults to False, and that a player which
  failed to open is never closed
- auth/loopback: assert the callback answers 200 and unknown paths answer 404

Remaining survivors here are equivalent/threading-internal (daemon flags,
join/wait timeout values, the sub-10Hz blocksize floor) and aren't observable.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

The human-table test asserted only the id/model columns, so blanking a present
created_at / audio_duration_sec value (the `value or ""` -> `and`) survived.
Assert those values appear in the rendered table.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

- account: assert the one-missing-bound window label, the one-day-window collapse,
  and that the default usage range spans exactly the last 30 days
- llm: assert `-o json` forces JSON output even for a non-agentic human
- login: assert the authenticated/logged_out success flags in --json output
- samples: assert the stream sample requests format_turns and that human-mode
  `samples list` renders its bullet list (pins the string concatenation)

(samples mkdir parents=True is equivalent here — the dir is one level under cwd.)

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

- assert the "Listening…" notice latches to exactly one emission
- assert begin/turn/termination are all forwarded to the renderer in non-follow
  mode (pins the follow-vs-None handler wiring)
- assert a turn event with no end_of_turn flag is treated as non-final in --llm
  follow mode (pins the getattr default)
- assert the renderer is closed even when streaming raises mid-run

Remaining survivors (worker-thread daemon flag, 0.1s join poll interval) are
equivalent/threading-internal and not observable.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

The new 200/404 status assertions need the response code. Use http.client
instead of urllib.request.urlopen so a 404 is a normal response status (not a
raised HTTPError), the status types as int (no mypy no-any-return), and no
urllib audit (S310) suppression — hence no new noqa escape hatch — is needed.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

The blocksize floor (max(1, rate//10)) and the rate>0 guard only differ for
sub-10Hz / 1Hz rates no real device reports — near-equivalent mutants, like the
daemon/timeout ones left elsewhere. Their fakes needed a `fake_sd: Any` module
(pyright can't assign attributes to a bare ModuleType), which tripped the
"no net-new Any" gate. The valuable resample-params and channels/dtype
assertions remain and add no Any.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

DEFAULT_ENV is already sandbox000, so the existing env tests couldn't tell the
`sandbox and env is None` override apart from the default. Bind a profile to
production and assert that a bare invocation keeps production while --sandbox
forces sandbox000 — killing the and/or and is/is-not mutants on that line.

(The two err=True echoes route to stderr; CliRunner mixes streams in this
version, so that routing isn't separately assertable here.)

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC

- init: assert a resolved key omits the skipped-key row, and that the
  launch-skipped row (with the manual uvicorn command) appears only when deps
  were installed AND no key is present — pinning both guards
- setup: a direct _proc_detail test pinning stderr-then-stdout preference

Remaining survivors in these modules (subprocess flag kwargs, timeout/exit-code
constants, the no-TTY picker guard) are low-value infra/near-equivalent.

https://claude.ai/code/session_014MYaBvEeWEJXJt36CCSHtC