feat(line): add opt-in smart reply modality

[jethac]

Summary

• Adds opt-in LINE smart reply modality routing behind smart_modality / LINE_SMART_MODALITY.
• Sends native voice replies for LINE voice/audio input and suppresses final text after successful voice delivery.
• Keeps text input as text and lets photo replies inherit the last text/voice modality for that LINE chat participant.
• Introduces the adapter reply delivery policy seam (reply_delivery_policy) that the modality routing rides on, with exception isolation and multiplex-safe adapter resolution.
• Carries regression tests for typed media-cache routing (_download_media dispatch), which main currently lacks.

Self-contained — one approval lands the whole former stack

This PR was previously presented as the top of a 3-PR stack (#35785 → #40931 → this). Both former base PRs are now closed in favor of this one: their entire diffs were already carried here as patch-equivalent commits, so the three open PRs presented ~triple review surface for one change. History on the review feedback:

• The sweeper review's re-scope of the media-cache change (test(line): regression coverage for typed media-cache routing #35785 thread) was implemented, and the adapter-side fix was then superseded by upstream 73e193c03 (fix(gateway): inbound voice classification/routing for Feishu, DingTalk, LINE, QQ, Photon, WhatsApp, Weixin #73515) — what survives here is exactly the regression coverage main still lacks.
• Both reply-delivery-policy review points (refactor(gateway): add adapter reply delivery policy #40931, 2026-07-14) are fixed on this branch with regression tests: adapter resolution via _adapter_for_source(event.source) (multiplex-safe), and the policy callback wrapped with a fall-back to the legacy voice gate so a policy failure can never disrupt final-reply delivery.
• Both smart-modality review points on this PR are fixed: streaming is gated pre-emission via a voice_reply_replaces_text pre-flight hook (with an integration test), and LINE_SMART_MODALITY is documented in both env-var tables.

Rebased onto current main (2026-08-21); the stale AUTHOR_MAP hunk was dropped (that dict is frozen — the contributor mapping now lives in contributors/emails/).

Related open work

#87397 proposes a per-platform suppress_text_when_voice config flag solving an overlapping problem in gateway/platforms/base.py. The policy seam here is the more general mechanism (per-adapter, exception-isolated, modality-aware); flagging so a maintainer can pick one mechanism deliberately rather than both landing.

Test Plan

• python -m py_compile gateway/run.py gateway/platforms/base.py plugins/platforms/line/adapter.py
• python -m pytest tests/gateway/test_reply_delivery_policy.py tests/gateway/test_line_smart_modality.py tests/gateway/test_line_plugin.py tests/gateway/test_platform_base.py -q (141 passed on the rebased head)

[jethac]

Rebased this feature PR onto the refreshed adapter policy branch (#40931).

Local verification:

uv run --with pytest --with pytest-asyncio --with aiohttp python -m pytest \
  tests/gateway/test_line_smart_modality.py \
  tests/gateway/test_line_plugin.py \
  tests/gateway/test_reply_delivery_policy.py \
  tests/gateway/test_voice_command.py::TestAutoVoiceReply \
  tests/gateway/test_media_download_retry.py \
  -q -o 'addopts='
# 138 passed

uv run --with pytest --with pytest-asyncio --with aiohttp python -m pytest \
  tests/gateway/test_voice_command.py \
  -q -o 'addopts='
# 163 passed, 21 skipped

[teknium1]

Thanks for the LINE modality work. The feature is still absent from current main, but two delivery/documentation issues need resolution before salvage.

Problems

• gateway/run.py:11401 suppresses only the normal final send. Current main marks already_sent after the final streamed content is confirmed delivered (gateway/run.py:20065-20073), and a None response means the adapter does not send another message (gateway/platforms/base.py:4894-4914). Thus, with streaming enabled, the voice reply can be followed by text that was already streamed.
• plugins/platforms/line/plugin.yaml:67 adds LINE_SMART_MODALITY, but neither website/docs/user-guide/messaging/line.md:159-177 nor website/docs/reference/environment-variables.md:549-569 documents it.

Suggested changes

• Gate/disable text streaming for smart voice-delivery turns before content is emitted, and add an integration test for the streaming-enabled path.
• Add the setting and its LINE_PUBLIC_URL audio-delivery prerequisite to both LINE documentation tables.

Automated hermes-sweeper review.

[jethac]

Thanks for the review — both points fixed; specifics in the inline replies, including two known limits of the streaming gate I'd rather disclose up front (interim-preview path, rapid-fire modality race). Rebased onto current main atop #40931. CI green, mergeable.

[jethac]

Consolidation update — this PR is now the single vehicle for the former #35785 → #40931 → #40933 stack.

• test(line): regression coverage for typed media-cache routing #35785 and refactor(gateway): add adapter reply delivery policy #40931 are closed in favor of this PR. Verified via git cherry: their entire diffs were already carried here as patch-equivalent commits, so the three open PRs were ~triple review surface for one change. One approval here lands everything.
• All review feedback across the stack is fixed and the threads are resolved — adapter resolution via _adapter_for_source (multiplex-safe) + exception-isolated policy callback with legacy-gate fallback (the refactor(gateway): add adapter reply delivery policy #40931 points), and the streaming pre-flight gate + LINE_SMART_MODALITY docs (this PR's points).
• Rebased onto current main today (head 69cf023405): all cherry-picks applied clean, the stale AUTHOR_MAP hunk is dropped (frozen dict; the mapping lives in contributors/emails/ now), and the touched-file suite is green locally (141 passed).

One thing worth a maintainer's deliberate pick: #87397 proposes a per-platform suppress_text_when_voice config flag that solves an overlapping problem in the same file. The policy seam here is the more general mechanism; happy to reconcile in either direction, but one of the two should be chosen before both bitrot.