fix(webchat): preserve coalesced replies and retry failed turns

[BegoniaHe]

Summary

WebChat coalesced fragments were finishing the pipeline with an empty completion (end) before the later flush event replied. Flush also reconstructed events through the four-argument constructor, so WebChatMessageEvent lost its queue manager and attachments directory.

Failed turns that wrote platform history but never a conversation checkpoint also could not be regenerated, edited, or used as a thread parent.

This keeps fragment events from emitting empty completion, clones the last event on flush, and lets regenerate/edit/thread continue from platform history when the conversation checkpoint is missing.

Rebased onto master after #7. No shared files remain.

Related issue

Related: #7

Root cause

• PipelineScheduler.execute always sent empty completion for requires_empty_completion after stages. TurnCoalesceStage stops the fragment and defers work to a later flush, but the stopped fragment still emitted WebChat end.
• _build_flush_event called last.__class__(message_str, message_obj, platform_meta, session_id). WebChatMessageEvent.__init__ also requires webchat_queue_manager and attachments_dir, so the flush event could not emit replies.
• ChatService regenerate, edit, and thread creation required find_turn_range to match a conversation checkpoint. A failed AI turn can have platform history plus llm_checkpoint_id without a matching conversation-history range.

Reproduction

1. Enable turn coalescing and send rapid consecutive WebChat messages. The first fragment emits end before the flush reply.
2. Fail a WebChat LLM turn, then regenerate, edit, or create a thread from that bot message. The API returns Linked checkpoint not found.

Implementation notes

• Mark coalesced fragments with event extra skip_empty_completion and honor it in the scheduler.
• Build the flush event with copy.copy(last) so adapter-specific fields stay attached, then reset per-run extras, result, and send flags.
• When turn_range is missing, skip conversation-history rewrite and continue from platform history for regenerate, edit, and thread creation. Older-turn branching checks still apply when a checkpoint range exists.
• Record the user-visible fixes in changelogs/v4.27.6.md.

Validation

uv run ruff format --check astrbot/core/pipeline/scheduler.py astrbot/core/pipeline/turn_coalesce/stage.py astrbot/core/pipeline/turn_window.py astrbot/dashboard/services/chat_service.py tests/unit/test_chat_service.py tests/unit/test_pipeline_scheduler_order.py tests/unit/test_turn_coalesce.py
uv run ruff check astrbot/core/pipeline/scheduler.py astrbot/core/pipeline/turn_coalesce/stage.py astrbot/core/pipeline/turn_window.py astrbot/dashboard/services/chat_service.py tests/unit/test_chat_service.py tests/unit/test_pipeline_scheduler_order.py tests/unit/test_turn_coalesce.py
uv run pytest tests/unit/test_pipeline_scheduler_order.py tests/unit/test_turn_coalesce.py tests/unit/test_chat_service.py tests/unit/test_delivery_receipt.py -q

79 passed after rebase onto master. Did not run make check or the full pytest suite.

Compatibility and risk

No OpenAPI, route, or public plugin-API change. skip_empty_completion is an event-local extra. Residual risk: other adapters that depended on constructor-only flush reconstruction now keep extra instance fields via copy.copy; WebChat is the adapter that required that. Failed-turn recovery does not rewrite conversation history when no checkpoint range exists. Older-turn branching checks are skipped when turn_range is missing.

Checklist

• The change is focused and does not include unrelated refactoring.
• I added or updated a regression test, or explained why a test is not practical.
• I ran the relevant formatting, lint, build, and test commands.
• User-visible behavior updates both docs/zh/ and docs/en/ when needed.
• OpenAPI, generated client, docs/public/openapi.json, and tests change together when routes or schemas change.
• No secrets committed. Runtime Python deps update pyproject.toml, requirements.txt, and uv.lock together.
• I did not restore legacy shims, Python <3.14 fallbacks, or upstream publish/docs URLs as fork artifacts.
• Breaking API or behavior changes use ! and a BREAKING CHANGE: footer.
• I will not merge this PR myself. Merge needs a human maintainer review plus a separate AI-assisted review (AI_POLICY.md).
• AI use follows AI_POLICY.md. Keep exactly one author note below. Do not fabricate the other.

Agent note

Goal: rebase #6 onto master after #7 so GitHub merge is clean, and record the WebChat fixes in changelogs/v4.27.6.md.

Paths: astrbot/core/pipeline/scheduler.py, turn_coalesce/stage.py, turn_window.py, astrbot/dashboard/services/chat_service.py, unit tests, changelogs/v4.27.6.md.

Checks run: Ruff format/check on the seven product files; uv run pytest on scheduler/coalesce/chat_service/delivery_receipt, 79 passed. Residual risk: full make check / make test not run here; failed-turn paths without a conversation checkpoint leave conversation history untouched.

Tools: grok-4.6 in opencode.

[BegoniaHe]

Split from #7

#7 dropped the two WebChat commits that duplicated this branch (skip empty completion and regenerate without checkpoint). Those patches were byte-identical; this PR still has the extra edit/thread follow-up e42855171.

This PR remains the sole owner of WebChat turn-handling. Merge with #7 in either order; no shared files remain.

I will not merge.

[BegoniaHe]

AI-assisted review (not a maintainer approval)

Separate from the authoring Agent note. GitHub rejects REQUEST_CHANGES on a self-authored PR, so this is a comment review. I will not merge.

Reviewed head cc0a16183 after rebase onto master (9a73e897e, #7). File overlap with the absorbed sync is gone. Focused pytest 79 passed in this session.

Verdict

Technically mergeable once CI on this head is green and a human maintainer approves. No blocking product defect found.

What looks correct

• Coalesced fragments set skip_empty_completion and stop_event(). Scheduler skips send(None), so WebChat does not emit end before flush.
• Flush uses copy.copy(last) then resets _extras, result, send flags, and temp-file/task sets. WebChatMessageEvent keeps webchat_queue_manager / attachments_dir. Flush does not inherit the skip extra.
• #9653 empty Respond skip is a different path (empty non-streaming chains). Fragments never reach Respond; flush carries merged content. They compose.
• Regenerating/editing/threading a failed turn without find_turn_range continues from platform history. When a range exists, older-turn branching checks still run.
• Tests cover skip-empty-completion, WebChat flush clone, and the three missing-checkpoint recovery paths.

Residual (non-blocking)

• copy.copy is shallow. WebChat needed the extra instance fields; other adapters now keep extra fields too.
• Missing turn_range skips conversation-history rewrite and older-turn checks. Intended for the last failed turn. An older failed bot message in a session that later continued can regenerate/edit/thread without the latest-turn guard.
• Changelog records the user-visible fix. WebUI docs already mention regenerate; no API/docs contract change.

Human maintainer still has to approve. This review does not authorize merge.