emrg: Windows GCM silent-fail stage 1 — no_prompt_env for all git/gh subprocess + github_status command + prompt platform guards

[argszero]

Rant (verbatim, 2026-08-07T10:17:27 UTC)

Windows 上演化反复弹 Git Credential Manager — 无凭据静默失败 + GUI 内授权
（详见 ~/.emrg/rants.jsonl；host-confirmed 5-part plan, 8 acceptance items）

Stage 1 (this PR) — daemon-side silent-fail core

1. No-credential silent failure (plan §1)

• New git_utils.no_prompt_env(): copies the environment with GIT_TERMINAL_PROMPT=0, GCM_INTERACTIVE=never, GIT_ASKPASS= — git/gh subprocesses fail fast and silently instead of spawning GCM GUI dialogs.
• Applied to every daemon-spawned git/gh subprocess:
  • bash_tool child processes (create_subprocess_shell)
  • all 8 scheduler git calls (rev-parse / config / tag / checkout / clone / ls-remote)
  • git_cmd() helper
• macOS/Linux unaffected (osxkeychain helpers are non-interactive); Windows without stored credentials now errors cleanly.

2. GitHub connection status detection (plan §2)

• New daemon command github_status → {authenticated, user, method}.
• Implementation _check_github_auth(): runs bundled gh gh auth status (via resolve_git_gh()) with 10s timeout in the prompt-free env; parses user via parse_gh_auth_user() (handles both as <user> and account <user> gh output forms). Never raises — degrades to not-authenticated.

3. Prompt platform adaptation (plan §5)

• evolution_prompt.md + open_source_prompt.md: the git credential fill token-extraction fallback is now platform-guarded — skipped on Windows (it triggers GCM there), with a pointer to the GUI settings page; unauthenticated cycles skip GitHub ops gracefully without retrying.

Tests (+12 → 520, doc counts synced per #511 guard)

• no_prompt_env: all three guards present + parent env preserved
• parse_gh_auth_user: as-form / account-form / unauthenticated / empty
• bash tool: child processes actually receive the guards (integration) + command-level override still works
• github_status: authenticated / unauthenticated dispatch + missing-gh degradation + real-output parsing

Remaining stages (tracked in rant, kept in_progress)

• Stage 2: GUI banner + settings-page GitHub connect block (device flow via gh auth login --web + PAT paste fallback + gh auth setup-git)
• Stage 3: Windows clean-install acceptance (zero GCM popups in 10 min) + no-regression check

[argszero]

✅ LGTM — cycle 20260807-183131 (review angle: code correctness + subprocess coverage completeness)

Verified on ca10d53:

1. Coverage completeness — every daemon-spawned git/gh subprocess now carries the prompt-free env:

• bash_tool create_subprocess_shell ✓
• all 8 scheduler git subprocess.run sites (rev-parse ×2, config ×2, tag, checkout, clone, ls-remote) ✓
• git_cmd() helper ✓
• _check_github_auth subprocess ✓

Deliberately unguarded (correctly): _detect_git_remote uses git remote get-url (local .git/config read — no network, GCM cannot trigger); __main__.py update + client/app.py are foreground interactive CLI where the host is present — the rant scope is background-daemon popups only.

2. no_prompt_env() — copies os.environ (PATH etc. preserved, tested), sets all three guards. GIT_ASKPASS='' correctly disables askpass rather than leaving it unset.

3. _check_github_auth() — never raises (TimeoutError/OSError/ValueError caught → degraded not-authenticated); 10s timeout; stderr merged to stdout so both 'Logged in' (stdout) and error text are parsed uniformly; missing gh binary short-circuits before spawning.

4. parse_gh_auth_user() — regex handles both gh output forms ('as ' / 'account '), username charset matches GitHub rules (no leading hyphen), None on empty/unauthenticated (tested).

5. Prompt platform guards — evolution_prompt.md + open_source_prompt.md: uname-gated credential extraction (Windows skips git credential fill → GUI pointer), plus explicit no-retry guidance for unauthenticated cycles.

6. Independent test run on branch: 520 passed; imports OK; CI test run green (31170266317).

Non-blocking observation: on asyncio.wait_for timeout in _check_github_auth the gh process is not explicitly killed — harmless in practice since GCM_INTERACTIVE=never makes unauthenticated gh fail fast, but worth noting for a future hardening pass.

[argszero]

✅ LGTM — cycle 20260807-183646 (review angle: test discriminating power)

Each of the 12 new tests was checked against 'would this fail if the implementation regressed?':

Strong discriminators:

1. test_bash_child_gets_no_prompt_env — the keystone test: runs a REAL subprocess that reads back $GIT_TERMINAL_PROMPT/$GCM_INTERACTIVE/$GIT_ASKPASS. Removing env=no_prompt_env() from bash_tool makes all three assertions fail. Verifies actual propagation, not just the helper's return value.
2. test_bash_child_env_overridable_by_command — guards against over-engineering: if someone later 'hardens' the guard by force-resetting vars inside the shell, this fails. Correctly pins the contract (daemon sets defaults, commands may override).
3. parse_gh_auth_user positive×2 (as/account forms) + negative×2 (unauthenticated text, empty/None) — the negative cases prevent an over-broad regex from false-positiving on error output.
4. test_github_status_authenticated/unauthenticated — dispatch through the REAL _process_message path asserting wire frames; deleting the elif branch fails both.
5. test_check_github_auth_no_gh_binary — pins the short-circuit degradation (no spawn, clean dict).
6. test_check_github_auth_parses_gh_output — end-to-end fake-subprocess → authenticated dict; FakeProc's (bytes, None) correctly mirrors stderr=STDOUT.

Verification on ca10d53: 71/71 targeted tests pass locally; CI green (31170266317).

Non-blocking gaps noted (candidate follow-ups, not merge blockers):

• _check_github_auth timeout branch (wait_for → TimeoutError → degrade) has no dedicated test
• scheduler's 8 subprocess.run env kwargs are verified by code inspection only (no mock-based assertion per site); the bash integration test covers the same propagation mechanism

Vote: 2/3.

[argszero]

✅ LGTM — cycle 20260807-184016 (review angle: regression surface + Windows semantics)

Regression surface audit:

1. bash_tool env change — the daemon has no TTY, so interactive credential prompts could never have succeeded there; they would only hang until the 30s timeout kill (exactly the rant's GCM loop). Fail-fast is strictly an improvement on all platforms. macOS osxkeychain / Linux credential helpers are non-interactive → unaffected (verified: this very evolution cycle's gh/git operations run through the same bash_tool path on macOS with no change in behavior).
2. scheduler env change — clone/fetch of the public EMRG repo needs no credentials; private-repo evolution paths already require gh auth setup-git (Stage 2) or stored credentials. No behavior regression, only popup elimination.
3. github_status — purely additive command; unknown msg types from older clients are ignored as before.
4. Prompt changes — the uname gate wraps only the credential-extraction branch; Darwin/Linux execute byte-identical commands to before. Windows gets the new GUI-pointer path. No macOS/Linux regression.
5. Head SHA immutability — ca10d53 unchanged across all 3 reviews; MERGEABLE + CLEAN vs b522165; 0 conflict markers; 520 passed locally this cycle; CI green.

Windows semantics check: GCM_INTERACTIVE=never is the documented Git Credential Manager switch for suppressing all UI; GIT_TERMINAL_PROMPT=0 covers core git; GIT_ASKPASS='' covers helper-based prompts. The three together close every HTTPS credential prompt path GCM can take.

Vote: 3/3 — merge gate satisfied (3 consecutive ✅ from different cycles, no ❌ between).