fix: report model failures returned as data and use single-token E2E markers

[AlexanderZ-Band]

Resolves INT-1641. Plan: https://linear.app/thenvoi/issue/INT-1641

Model failures returned as data

Agno, Google ADK and Gemini return a failed model run as a value instead of raising, so a blocked call could end a turn silently. If the turn had already done real work, it was even marked complete.

• ProviderRunError(code, detail) (band.core.exceptions) and generic_provider_failure(provider, error) (band.core.protocols) are the one rule. The room sees GENERIC_PROVIDER_FAILURE_MESSAGE plus the coarse code. detail (provider text) only goes to the log.
• Agno: AgnoRunError now subclasses ProviderRunError. Behavior is unchanged.
• Google ADK: a non-STOP event.error_code is recorded and raised once the run drains. Draining avoids leaving ADK's generators open. ADK ≤1.10 reports an empty normal finish as "STOP", which stays a normal end of turn.
• Gemini: _model_failure(response) reads the response the way ADK's LlmResponse.create does. A reply with content (MAX_TOKENS included) or an empty STOP is not a failure.
• docs/turn-outcome.md documents the rule.

E2E markers

• unique_marker(label) returns one uppercase token (NOTE0A832EC4) and rejects non-alphanumeric labels. Models dropped the word-like note- prefix.
• Hyphenated labels are renamed. The prefix coaching in the cross-framework rehydration recall is removed, and test_copilot_sdk uses unique_marker instead of inline markers.

Diagnostic

• test_invites_messages_and_removes_a_peer now prints the agent's tool calls and failed results when the add is missing.

Tests

• ADK runs a real InMemoryRunner against a scripted local BaseLlm registered in LLMRegistry. Responses go through ADK's own LlmResponse.create.
• Gemini gets real GenerateContentResponses at the client boundary.
• Both judge turns through on_event. Cases covered:
  • SAFETY, a blocked prompt, and (ADK only) no candidates
  • SAFETY after real work
  • empty STOP gives exactly the missing-reply failure
  • Gemini: MAX_TOKENS after a reply completes
  • usage is still emitted on failure
• Every regression case fails with its adapter fix reverted. Verified on ADK 1.1.1 (py3.14) and 1.10.0 (py3.12).
• Full unit suite, ruff, pyrefly and the markdown docs pass.

Departures from the plan: the "real work" tool in the tests is band_create_chatroom, which is advertised by default (the plan named band_create_task, which needs the tasks capability). Gemini's check returns the error itself rather than only a code, so the provider message stays attached for the log.

Live E2E

All runs are on c03d192 with first_attempt_diagnostics=true (no retry rescue).

Run	Result
core, test_rehydration_cross_framework or test_participant_management, 1/3	✅ 37437871435
same, 2/3	❌ 37438460290: test_rehydrates_foreign_peer_message[pydantic_ai], core / ubuntu. The recall turn ended with no message. It failed the same way on main with the old prompt (37094178741), so the marker change didn't cause it. Tracked in INT-1698.
same, 3/3	✅ 37439206975
core, full	✅ 37439751401
google, full	✅ 37437855994
backends, full	❌ 37437852493: ubuntu ✅. Windows failed only test_usage_not_cumulative_across_turns[opencode], which timed out mid-turn. This PR doesn't touch that cell, and it has failed on main before (36517841221, 35558057110, 35486844106). Tracked in INT-1699.

core runs were dispatched one at a time, because a newer core run on the same branch and OS cancels the one in progress.

🤖 Generated with Claude Code

https://claude.ai/code/session_017jhEdA7VR9ng3nZpKW2Z2X

[linear-code]

INT-1641

[amit-gazal-band]

Approving. One small fix worth making before merge:

Gemini _model_failure misses ADK's no-candidates case (src/band/adapters/gemini.py, _model_failure)

The docstring says this reads the response the way google-adk's LlmResponse.create does. For a response with candidates=[] and no prompt_feedback, ADK returns error_code='UNKNOWN_ERROR', but _model_failure returns None. So the ADK adapter fails that turn (as test_no_candidates_fails_with_adks_own_code checks), while the Gemini adapter treats it as a normal finish. If the turn already did real work, for example a tool round, it gets judged completed and the failure is lost. That is the silent-failure case this PR is meant to fix.

Suggested fix: in the no-candidates branch, return ProviderRunError("UNKNOWN_ERROR") when there is no block_reason, and add the matching Gemini test case (the shared Google model-failure cases already include no-candidates for ADK).