emrg: daemon observability (exit records + crash log) & promote blog cadence 1-3 days

[pm25coder]

Two independent fixes driven by two host rants.

1. Daemon observability — durable exit records + stderr crash log

Rant (verbatim, 2026-08-25T09:25:32+08:00):

daemon 异常退出无任何日志痕迹，死因无法定位（已 2 例可考：8-24 20:40 腰斩 R46、8-25 02:00 腰斩 emrg-task 01:26 轮并跳过 promote 06:06 调度）。emrgd.log 在死亡瞬间无 shutdown/错误/退出码记录，重启后一切正常。请求日志/可观测性增强（非防死）：① daemon 退出无论正常/异常都写退出记录（退出码、信号、traceback）；② stderr/stdout 独立重定向到崩溃日志文件；③ TaskHandler 运行中 cycle 定期持久化进度心跳，重启后能识别"上次中断于 round X"；④ scheduler next run 持久化，重启不重置。目标：下次再出现时能定位原因。

Root cause: daemon_manager spawns emrgd with stderr=DEVNULL, so anything bypassing logging — asyncio's default exception handler, C-level abort traces, faulthandler dumps — vanished without a trace. A background-task crash or fatal signal left zero evidence in emrgd.log.

This PR covers acceptance items ① and ②:

• DaemonExit + _write_exit_record (daemon.py): every stop path of run_server (normal return, serve() crash, SIGINT, SIGTERM, cancellation, unexpected BaseException) now produces a DaemonExit(reason, exit_code, traceback) that the entry point persists as one-line JSON in ~/.emrg/emrgd-exit.log (append-only, survives emrgd.log rotation) and mirrors into emrgd.log. Exit-code 0 vs 1 vs 130/143 distinguishes clean stops from crashes.
• stderr/stdout redirect (__main__.py): _redirect_std_streams() points sys.stdout/sys.stderr at ~/.emrg/emrgd-crash.log (independent sink, keeps the file handler on emrgd.log separate) and enables faulthandler so even fatal-signal deaths (SIGSEGV, no Python code runs) leave a stack dump. StreamHandler is now only attached for interactive TTY runs, so it cannot duplicate the log into the crash file.
• _asyncio_exception_handler (daemon.py): installed in run_server, routes background-task crashes (TaskHandler, websockets callbacks) into emrgd.log with the task identity + traceback instead of the default handler's stderr → DEVNULL void.
• _sigterm_handler: POSIX SIGTERM → SystemExit so SIGTERM stops are attributable (Windows keeps its hard-kill path, guarded).
• Tests (+2, Agent.md count 1052 → 1054): test_write_exit_record (durable JSON records) and test_asyncio_exception_handler_routes_to_logger (background crash reaches the logger).

Items ③ (TaskHandler progress heartbeat) and ④ (scheduler next-run persistence) are staged for a follow-up PR — this change already makes every daemon death attributable.

2. Promote prompt — blog cadence 1-3 days per post

Rant (verbatim, 2026-08-25T10:01:20.923337+08:00):

promote_prompt.md blog 发布节奏过慢：≤1 篇/周 → 应改为 1-3 天一篇

§2.y Blog Publishing 与 channel 表写死 low frequency, high quality — default ≤1 post/week。当前背景下该节奏不匹配：① 已有 5 篇 publish-ready 草稿积压，r15 完成的 postmortem（draft 5）因周更节奏硬等一周，窗口推到 08-27；② 项目 8 releases/2 天的迭代速度使草稿素材在等待中过期，需反复刷新版本号续命；③ 草稿选题彼此独立（postmortem / git-as-state / 上下文预算 / fail-loud / 8-releases），无重复主题风险。

请将节奏改为：每 1-3 天发布一篇（至少 2 篇/周），同时保留现有质量门槛——深度 > 长度、发布前按 §0.4 fact-check 到最新 commit/release、非硬广。并同步更新 channel 表的 low cadence (≤1 post/week) 表述，避免两处规则打架。

• promote_prompt.md §2.y Blog Publishing: cadence rule changed from low frequency, high quality — default ≤1 post/week to 1-3 days per post (at least 2 posts/week), keeping the quality gates (depth > length, §0.4 fact-check, no hard ads).
• Channel table row for Blogs: low cadence (≤1 post/week) → cadence 1-3 days per post (see Blog Publishing) so the two spots no longer contradict each other.

Verification

• uv run pytest tests/ -q → 989 passed, 65 skipped (1054 collected, doc-count guard updated)
• Import check from emrg.client.app import run_client OK; python -m emrg --help OK

[argszero]

✅ LGTM — cycle

Reviewed the full diff (daemon observability + blog cadence, rant 2026-08-25T09:25:32 / 10:01:20):

• emrg/server/__main__.py: StreamHandler gated on sys.stderr.isatty() (daemon_manager spawns with stderr=DEVNULL — the handler was pure waste and would duplicate the log into the crash log after redirect); _redirect_std_streams() best-effort (OSError → keep DEVNULL) + faulthandler enabled with the stream; main() wraps asyncio.run and records DaemonExit on EVERY path (sigint/sigterm/crash/normal) with sys.exit(result.exit_code).
• emrg/server/daemon.py: DaemonExit (slots) → _write_exit_record appends one JSON line to ~/.emrg/emrgd-exit.log (survives emrgd.log rotation) + mirrors into emrgd.log; _asyncio_exception_handler routes background-task crashes to the RotatingFileHandler instead of stderr=DEVNULL (task identity + exc_info kept); _sigterm_handler POSIX-only (guarded for Windows); serve() crash now captures _crash_traceback. run_server returns DaemonExit on normal/cancel/sigint/BaseException paths — every stop is attributable.
• promote_prompt.md: blog cadence ≤1/week → 1-3 days (at least 2/week) with rationale; consistent between the channel table and the Blog Publishing section.
• Tests: test_write_exit_record (durable JSON, traceback/timestamp/pid) + test_asyncio_exception_handler_routes_to_logger (exc_text/ValueError verified). Agent.md 1052→1054 matches the +2 tests (#511 guard green on both jobs).
• CI: test + test-windows both green (run 32801450030).

No issues found; solid follow-through on the silent-death observability rant.

[argszero]

✅ LGTM — cycle (re-review)

Head unchanged (bba65b5), CI double-green (run 32801450030). Re-verified against the previous review's findings: DaemonExit write path (survives emrgd.log rotation), _asyncio_exception_handler routes background-task crashes to the RotatingFileHandler (stderr=DEVNULL otherwise swallows them), _sigterm_handler POSIX-guarded, all run_server exit paths return a DaemonExit with the exit code, and the blog cadence change (≤1/week → 1-3 days) is consistent between the channel table and the Blog Publishing section. No issues found — 2/3.

[argszero]

✅ LGTM — cycle (3/3)

Head unchanged (bba65b5), CI double-green (run 32801450030), MERGEABLE. Full diff was reviewed in the two prior cycles (daemon observability: DaemonExit exit records, crash-log redirect via _redirect_std_streams + faulthandler, asyncio exception handler routing to RotatingFileHandler, POSIX-guarded SIGTERM attribution, run_server DaemonExit on every exit path; blog cadence 1-3 days consistent in channel table + Blog Publishing section; Agent.md doc count 1052→1054 matches the +2 tests). No issues found — merging.