fix(backend): treat any non-end_turn stop as a failed side-channel call

[philmerrell]

What

The side-channel Converse calls only rejected stopReason == "max_tokens". Any other non-completion stop (a guardrail stop, a content-filter refusal, tool_use, an unknown reason, or a missing reason) was accepted as a valid result. This change accepts a result only on a positive completion signal. Everything else goes through the fallback that site already has. No new fallbacks are added.

Site	Accepted	Anything else goes to
Compaction narrative (compaction_summary._compress)	end_turn, max_tokens (salvaged as before)	None, then newest-first truncation (truncated_after_model / extract_then_truncate)
Compaction extraction (extract_with_model)	end_turn, max_tokens (complete lines kept as before)	None, then the narrative alone (plain compression)
Tool-batch summary (tool_summaries/summarizer.py)	end_turn	None, then the SPA's deterministic formatter
Document abstract (files/document_digest.py)	end_turn	None (digest without an abstract)
Conversation title (inference_api/chat/service.py)	end_turn, max_tokens (the existing 50-char clip)	The "New Conversation" placeholder, with no title write and no session_title event

Why

Kaizen review 2026-09-25, Proposal 2 (Ship). A refused or guardrail-stopped compaction summary becomes the checkpoint in the cacheable conversation history, and it stays there until the next cut. That costs answer quality for the rest of the session and cannot be undone. The tool-summary, abstract and title outputs also persist (as TSUM# rows, the digest and the session title), so a refusal's text would be shown or fed back to the model. We don't know exactly which stop reason Bedrock returns for a refusal, and this change makes that irrelevant.

The max_tokens salvage added for compaction in PR 1353 is unchanged. The title keeps accepting max_tokens because its overrun handling (the 50-char clip) already exists and changing it is out of scope.

Tests

• One stubbed guardrail_intervened test per site, parametrized with content_filtered, tool_use, an unknown reason and None where relevant:
  • Compaction: narrative refused on a single call, extraction refused, and narrative refused on the extract path.
  • Tool summary, document abstract and title: the refused result is dropped.
• The existing max_tokens tests are unchanged and still pass, including all salvage tests. Added a title test confirming max_tokens output is still clipped.
• Touched files (plus the quality harness and the title route tests): 258 passed.
• Full backend suite: 10890 passed, 3 skipped, 0 failed.

🤖 Generated with Claude Code