fix: Python 3.14 compatibility (daemon_pool worker signature, gateway liveness argv, dotenv reload race)

[slow4cyl]

Three small fixes that together make the agent run correctly on Python 3.14. Each is its own commit:

---

fix(daemon_pool): Python 3.14 ThreadPoolExecutor._worker signature compat

Python 3.14 changed ThreadPoolExecutor._worker to (executor_ref, ctx,
work_queue), folding initializer/initargs into a worker context object.
The subclass still passed the old 4-tuple, so every pool thread died at
spawn: on 3.14 the kanban dispatcher spawned zero workers. Branch on the
presence of _create_worker_context so 3.8-3.13 keep the old path.

fix(gateway): status liveness check tolerates setproctitle-stripped argv

main.py calls setproctitle("hermes"), erasing the gateway subcommand
from /proc cmdline. The liveness check then failed its "looks like a
gateway" test and reported a healthy systemd gateway as OFFLINE. When
argv carries no gateway subcommand but PID + start_time match the
persisted record, defer to the recorded gateway identity.

fix(env_loader): retry dotenv reload on transient KeyError

python-dotenv's env.update(os.environ) can race another thread mutating
os.environ (the in-gateway kanban decomposer pins/unpins env vars per
board), raising a transient KeyError that crashed in-gateway cron jobs.
Retry once on that specific failure.

---

🤖 Generated with Claude Code

[alt-glitch]

This was generated by AI during triage.

Related, not a duplicate. The tools/daemon_pool.py half of this omnibus overlaps the earliest-open Python 3.14 _worker-signature fix #57459 (and its closed siblings), but this PR bundles two additional novel 3.14 fixes not covered there: the gateway/status.py liveness check tolerating setproctitle-stripped argv, and the hermes_cli/env_loader.py dotenv-reload retry on transient KeyError. Flagging the daemon_pool overlap so a reviewer can decide whether to take the daemon_pool half from #57459 and the other two from here, or land this whole PR.

[teknium1]

Thanks for the focused compatibility work. The daemon-pool branch matches CPython 3.14's _create_worker_context() / three-argument _worker shape, and current main still has the pre-3.14 call in tools/daemon_pool.py:52-60.

Problems

• The PR cannot yet make Hermes runnable as a supported Python 3.14 installation: pyproject.toml:13-20 deliberately caps requires-python below 3.14 because required Rust-backed dependencies lack cp314 wheels.
• hermes_cli/env_loader.py:162 invokes the Latin-1 fallback outside the new KeyError retry handler. A transient KeyError in that fallback escapes rather than retrying.
• No regression tests are included for the three changed paths; relevant suites already exist at tests/tools/test_daemon_pool.py, tests/gateway/test_status.py, and tests/hermes_cli/test_env_loader.py.

Suggested changes

• Put both encoding paths inside the KeyError retry loop, then add focused regressions for the 3.14 worker signature, bare hermes gateway argv, and KeyError on both encoding paths.
• Re-scope the Python 3.14 claim or separately resolve the packaging constraint in pyproject.toml:13-20.

Automated hermes-sweeper review.

[teknium1]

This Latin-1 fallback is outside the new except KeyError block. If the same concurrent-environment KeyError occurs after UTF-8 decoding fails, it escapes without a retry; wrap the encoding fallback inside the retry attempt.

[slow4cyl]

Re-scoped as requested: this is forward-compatibility for 3.14 — every 3.14 branch is feature-detected and a no-op on ≤3.13, and requires-python was never bumped here (stays capped until the Rust-backed deps ship cp314 wheels). env_loader now runs both encoding paths inside the KeyError retry loop, with a regression pinning the latin-1-fallback case you flagged. While adding the requested status tests I found the record-fallback was too broad — a readable s6-supervise gateway-coder cmdline on a matching-start-time PID regressed test_runtime_status_running_pid_rejects_stale_record_for_supervisor_pid — so it's narrowed to exactly the stripped setproctitle("hermes") shape (_cmdline_is_bare_hermes_proctitle); all other readable cmdlines keep strict judgment. Added regressions in all three suites you named, including stand-in executors that pin both _worker arg shapes on any interpreter and an interpreter-probe guard so a future stdlib reshape fails loudly. On a 3.14 interpreter the four pre-existing daemon_pool tests fail on main and pass here; no other failure-set changes. Rebased onto current main.