docs(synthesizer): document + version + test synthesizer behavior (v1.0 #5)

[ernestprovo23]

PR-C of conclave v1.0 — document + test synthesizer behavior

Closes readiness must-do #5 (~/.claude/sdlc/conclave-v1/00_READINESS_REVIEW.md): the synthesizer path is the heart of the "council" value prop but was undocumented and lightly tested, risking silent degradation.

Investigation — actual current behavior (file:line)

• Which model synthesizes: Council.synthesizer = synthesizer or config.synthesizer (council.py:76); CLI --synthesizer/-s (cli.py:245); config synthesizer: (config.py:80/191); default claude = anthropic/claude-sonnet-4-6 (registry.py:138). Same model is the judge (adversarial, modes.py:421) and debate consolidator (modes.py:237).
• Failure / unkeyed path: already OBSERVABLE, not silent. No-usable-answers (council.py:364), synthesizer unkeyed (council.py:372), synthesizer call fails (council.py:390) each set CouncilResult.synthesis_error; synthesis stays None; member answers preserved. Adversarial judge mirrors to AdversarialResult.verdict_error → synthesis_error (modes.py:431/454/344). No silent quiet-concat path existed — confirmed + pinned with tests rather than "fixed."
• Prompt: _SYNTH_SYSTEM constant (council.py), reused by streaming (streaming.py:189); debate/judge prompts in prompts.py. Was not versioned → addressed here.

What changed

1. Versioned synthesis prompt — conclave.prompts.SYNTHESIS_PROMPT_VERSION ("2026-06-14"), re-exported from council, stamped onto every CouncilResult as prompt_version (lazy default_factory avoids the prompts↔models import cycle). Prompt text byte-stable; constant + text pinned so a prompt change without a version bump fails CI.
2. Documentation — module + _synthesize docstrings (selection precedence, the three degraded paths, versioning) and a README "Synthesizer behavior" section. Doc index: new test row + CouncilResult field note + changelog row.
3. Confirmed observability — no behavior change to the degraded path; regression tests assert the signal is present.
4. tests/test_synthesizer.py (21 tests) — (a) default selection, (b) arg/config override, (c) CLI --synthesizer override, (d) degraded path signaled for synthesize + debate + adversarial judge (unkeyed + call-failure), (e) prompt-version stability across every mode. Mocks at the existing httpx-transport boundary (offline).

Invariants

• No non-synthesis behavior changed; happy-path synthesis output is byte-for-byte unchanged (additive prompt_version metadata only).
• ruff check + ruff format --check clean; coverage 89.39% (CI floor 75%); 210 passed.
• The 2 test_logging.py failures seen locally are a pre-existing caplog handler-isolation artifact (byte-identical to main, green in CI), unrelated to this PR.

🤖 Generated with Claude Code

[ernestprovo23]

Added a second commit (e0cf138): CI on this branch resolved pytest 9.1.0 (deps are >=8.0.0; main's last green run predates 9.1.0). pytest 9.x attaches its LogCaptureHandler (a StreamHandler subclass) to the non-propagating conclave logger, so test_logging.py's len(handlers) == 1 assertions saw 3 handlers and failed across all 3 Python versions. Fix counts only conclave's own handler via type(h) is logging.StreamHandler (pytest's is a subclass) — intent-preserving, no production code changed. This is a pre-existing latent fragility surfaced (not caused) by the pytest upgrade.

[ernestprovo23]

Superseded by #45 (integrated into the v1.0.0 release commit).