refactor(llm,vision): canonical codec_llm + codec_vision helpers (A-11 + A-12 tranche 1)

[AVADSA25]

Summary

PR-3E (Option 2). Two new single-source modules replace hand-rolled duplicates on the repo's hottest path (every feature calls an LLM / vision). Design-first per AGENTS.md §11 → docs/PR3E-LLM-VISION-DEDUP-DESIGN.md (now flipped to IMPLEMENTED, see §8).

A-11 — vision dedup (fully closed)

• New codec_vision.py: describe_sync + describe_async, Gemini-flash (gemini-2.0-flash) → local-Qwen-VL fallback, config read live from codec_config (provider/model/Keychain changes take effect without restart).
• All 3 consumers delegate:
  • codec.py vision_describe → describe_sync (deleted _gemini_vision / _local_vision).
  • codec_voice._analyze_screenshot → await describe_async(..., http=self._http) (reuses the pipeline's httpx client).
  • codec_session.screenshot_ctx → describe_sync — gains the Gemini fallback it previously lacked (documented behavioral superset).

A-12 — chat/completions caller (first tranche)

• The audit's premise that codec_llm_proxy already had call()/stream() was inaccurate — that module is a priority queue (semaphore), not an HTTP caller.
• New codec_llm.py (config-agnostic, no import cycle): call() + strip_think / extract_content — headers, Bearer auth, chat_template_kwargs.enable_thinking, <think> strip, choices/reasoning parse, retry + 2**n backoff, never raises.
• Migrated codec.py voice-reply chat + codec_session.qwen_call; removed the now-dead local extract_content in codec_session (canonical copy lives in codec_llm).

Deferred (each its own design + PR)

• codec_session.qwen_stream SSE → needs a codec_llm.stream() generator.
• The remaining ~40 chat/completions sites (dashboard, voice generate_response, agents/agent_plan/agent_runner, telegram/imessage bridges, compaction/self_improve/watcher/textassist/dictate).

Notes / behavior deltas (documented)

• codec.py voice path: non-200 and empty responses now collapse to one apology ("Sorry, I didn't get a response.") instead of two distinct messages.
• codec_llm.call: on 200-but-empty it returns "" without retrying (codec.py parity exact at retries=1; for qwen_call retries=3 this skips pointless identical retries — an improvement).
• Net -86 LOC in tracked files (+3 new modules/tests).

Test plan

• tests/test_llm_vision_dedup.py — 19 tests: strip_think/extract_content matrix; call success / no-key-omits-auth / retries-then-empty / exception-empty; describe_sync gemini-first / fallback / local-only / both-fail; describe_async gemini + fallback (driven via asyncio.run + fake httpx client — no pytest-asyncio dependency); source-invariant checks (codec.py/voice/session call the canonical helpers, inline impls gone).
• Full suite: 23 known-baseline failures, zero new (1395 passed, 74 skipped).
• Ruff: new modules + test clean; codec.py unchanged baseline (E402 from its signal-before-imports pattern); no new F401.
• No skills/ touched → no manifest regen.
• Manual (Mac Studio): voice "look at my screen" + a chat screenshot describe correctly via both providers; a voice chat turn replies.

🤖 Generated with Claude Code