fix(gateway): stop lease-wait refreshes flooding non-editing chat adapters

[liuhao1024]

What does this PR do?

While a durable session's cross-process turn lease is held by another Hermes process, the waiting process emits an initial "Another Hermes process is using this session…" notice once, then a periodic "Still waiting… (Ns)" refresh roughly every 15s. That refresh only makes sense on surfaces that can update the initial notice in place. On adapters without send_or_update_status (WeCom, Weixin, QQ, Signal, … — all SUPPORTS_MESSAGE_EDITING = False) the status path falls back to a plain adapter.send, so every refresh landed as another standalone chat message: a two-minute wait produced eight separate messages that drowned the eventual delivery.

This PR suppresses the periodic refresh when the adapter cannot update the status in place:

• Extracts the refresh wording into a module-level template constant at the emit site.
• Derives a compile-once matcher from that constant (same constants-are-the-wording convention as _COMPRESSION_PROGRESS_STATUS_RE, fix(gateway): suppress routine pre-API compression chatter on chat platforms #69550) and gates it on adapter capability in the status callback.
• Adapters that CAN update in place (Telegram, Slack) keep refreshing the existing bubble unchanged.

The initial wait notice and the lease-timeout warning use different wording and are always delivered, on every platform.

Related Issue

Fixes #89166

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature
• Breaking change
• Documentation update

Changes Made

• SESSION_TURN_LEASE_WAIT_REFRESH_STATUS_TEMPLATE: the periodic refresh wording as a format template; the emit site now formats it instead of an inline f-string.
• _session_turn_lease_refresh_re(): lazy compile-once regex derived from that template via the existing _status_template_to_regex helper (lazy because this module never imports run_agent at module scope).
• _should_suppress_lease_wait_refresh(adapter, message): True only when the adapter has no callable send_or_update_status AND the message is the periodic refresh.
• _status_callback_sync: early-returns (with a debug log) when that helper says the refresh would flood this adapter.
• New regression test suite covering: refresh suppressed across elapsed values on a no-update adapter, refresh delivered on an update-capable adapter, initial notice / timeout warning / unrelated lifecycle statuses never suppressed, and the refresh surviving the noise filter before the gate runs.

Related but distinct from #12610 (which suppresses ALL lifecycle events on the email platform by platform enum): this change is scoped to the single periodic lease-wait refresh template and keys off adapter capability, so every other lifecycle status keeps flowing to non-editing platforms.

How to Test

python -m pytest tests/gateway/test_session_lease_wait_refresh.py -q
# 5 passed

python -m pytest tests/gateway/test_telegram_noise_filter.py tests/gateway/test_compression_progress_notices.py tests/gateway/test_telegram_status_update.py tests/state/test_session_turn_lease.py tests/run_agent/test_cross_process_turn_lease.py -q
# all passed locally

Manual repro from the issue: open the same durable session from Desktop and send a DM through a WeCom/Weixin gateway while Desktop holds the turn; previously each 15s tick posted a new chat message, now only the single initial wait notice appears (and the lease-timeout warning if the wait exceeds the budget).

Fail-on-main check: with the source changes stashed, the new suite fails at collection (the gate and template constant do not exist on main).

Checklist

• Code follows the project's style guidelines
• Self-review completed
• Comments explain intent (why the refresh must be capability-gated)
• Corresponding changes documented via docstrings
• Tests added and passing (5 new; adjacent suites green)
• Single root cause, no unrelated changes

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Reviewed by reviewer-e (AI automated review).

Clean fix for notification flooding with the right single-source discipline: the refresh wording moves into SESSION_TURN_LEASE_WAIT_REFRESH_STATUS_TEMPLATE and the gateway derives its matcher from that same constant (never re-inlining the prose — explicitly called out as the #69550 convention), so emitter and suppressor can't drift. Suppression keys on real adapter capability (send_or_update_status present) rather than a platform name list, the initial notice and lease-timeout warning are provably immune by wording difference, and the five tests cover suppressed/delivered/initial/timeout/unrelated shapes including a "refresh survives the noise filter first" assertion.