fix(agents): salvage experiment answer when weak model fails structured output

[w7-mgfcode]

Summary

Follow-up to #347/#348 and #349/#350. The Experiment Agent still fails read-only queries on local Ollama models, but the true root cause (captured via a full capture_run_messages trace) is not a tool-loop the model chooses — it is structured-output (PromptedOutput(ExperimentReport)) incompatibility with weak 8B models, which no prompt change can fix.

Trace evidence (live model ollama:llama3.1:8b)

msg[1] TOOLCALL tool_list_runs {"status":"success"}   ← real tool call
msg[2] TOOLRETURN → 10 runs incl. WAPE                ← data obtained
msg[3] TEXT: {"runs":[...],"total":10,"page":1}       ← model returns RAW tool data as its "answer"
msg[4] RETRY: summary: Field required                 ← not an ExperimentReport
msg[5] TOOLCALL tool_list_runs {} ... loops 3x → "Exceeded maximum output retries (3)" → UnexpectedModelBehavior

• The model calls the tool and gets the data, but emits the raw tool-result shape ({"runs":[...]}) as its final output instead of ExperimentReport{summary}. PromptedOutput rejects it (missing summary), and it loops to retry-exhaustion.
• qwen3:8b is worse: it emits the tool call itself as text ({"tool":"tool_list_runs","arguments":{...}}) and never makes a real call.
• Both prompt-based mitigations (fix(agents): constrain experiment read-only queries #348 guard, fix(agents): stop experiment read-only tool-call loop on weak models #349 one-pass rule) cannot fix this — the model isn't choosing to loop; it can't produce the required structured output while juggling tools.

Fix

Service-layer graceful finalizer fallback: when an agent run raises UnexpectedModelBehavior ("Exceeded maximum output retries") and there is no pending HITL action to salvage, extract the tool results already obtained during the run (via capture_run_messages) and make ONE tool-less, plain-text (output_type=str) follow-up call to the same model to answer the user's question from that data. Plain text is what weak models can produce; this converts the structured-output failure into a correct answer, on Ollama, without a cloud model.

• Keeps PromptedOutput(ExperimentReport) as the happy path (cloud models unaffected).
• No tools on the finalizer → cannot loop; str output → cannot fail schema validation.
• HITL approval salvage (fix(agents): Ollama chat HITL — sanitize null content and preserve pending approval #344) takes precedence and is untouched.

Acceptance

• Finalizer fallback wired into both chat() and stream_chat() misbehavior handlers.
• Deterministic unit tests for tool-payload extraction + handler behavior (returns salvaged answer when data exists; generic error when none). No live model calls in tests.

[w7-mgfcode]

Fixed in PR #352 (service-layer plain-text finalizer fallback + tool-payload compaction), merged to dev. Live-verified on ollama:llama3.1:8b.