refactor(llm): codec_llm.stream() + migrate streaming/trivial sites (A-12 tranche 2)

[AVADSA25]

Summary

PR-3E-2 (Option 1). Builds the streaming keystone codec_llm.stream() and migrates the lowest-risk remaining chat/completions sites. Design-first per AGENTS.md §11 → docs/PR3E2-LLM-STREAM-TRANCHE2-DESIGN.md (maps all 22 remaining sites + the phased roadmap).

New API

• codec_llm.stream(...) — sync SSE generator yielding the RAW content deltas (data: framing, [DONE] sentinel, choices[0].delta.content, per-chunk parse tolerance). Never raises — HTTP/conn/parse error → empty stream, caller falls back. Think-stripping is intentionally left to the caller (see note below).
• _build_request() — shared header/auth/payload builder so call() and stream() can't drift; stream flag applied last so extra_kwargs can't clobber it. call() refactored onto it with a byte-identical payload (no stream key) — PR-3E's call() tests stay green.

Migrated

• codec_session.qwen_stream (streaming proof) — keeps live stdout + strip_think() on the accumulated result; falls back to the already-migrated qwen_call on an empty stream.
• codec_compaction.compact_context — single httpx POST → call(); also gains enable_thinking=False + <think> strip (slightly cleaner summaries), fallback unchanged.
• codec_dictate draft-refine — hardcoded :8083 now passed as base_url; all failure branches collapse to "use raw body", which never-raise→"" maps onto exactly.

Why stream() yields raw (not think-stripped)

qwen_stream writes each delta to stdout live and only strip_thinks the accumulated result — internal stripping would silently drop the reasoning it shows live (a parity break) and needs fiddly partial-tag buffering. Raw-yield gives exact parity, is simpler/safer for the first streaming API, and matches reality: the dashboard (deferred) owns its own cross-chunk <think> + [SKILL:…] machine.

Scope refinement (read-the-source)

The approved Option 1 listed 4 trivials; reading the downstreams moved 2 to tranche 2c:

• codec_textassist RAISEs on failure → its except shows an Error overlay. With never-raise→"", the success path would pbcopy "" + ⌘V, pasting empty over the user's selection + "Text replaced!". Destructive — needs raise_on_error.
• scripts/regen_skill_descriptions uses raise_for_status() (fail-loud dev script) — never-raise would write empty descriptions.

Both share the raise-on-failure contract with agent_plan/agent_runner, so they migrate together in 2c when codec_llm.call gains raise_on_error. (Honors "never break working code.")

Still pending (own PRs)

2c raise-mode (textassist/regen/agent_plan/agent_runner) · async astream() for voice _stream_qwen + agents (queue stays at the call site) · dashboard (4 non-stream + the [SKILL:…] stream tag-machine) · bridges · skills tranche.

Test plan

• tests/test_llm_stream.py — 14 tests: stream() raw-deltas/[DONE]/blank+garbage/non-200/exception/auth/payload-parity-with-call(); qwen_stream consume+strip_think + fallback-on-empty; compaction use + fallback; source invariants (session/compaction/dictate use the canonical helpers, inline impls gone).
• Full suite: 1409 passed, 23 known-baseline failures, zero new, 74 skipped.
• Ruff: new code 0 errors; codec_dictate 24-before/24-after (all pre-existing E701), codec_session 1 pre-existing E741 — zero net-new.
• No skills/ touched → no manifest regen.
• Manual (Mac Studio): a streaming session reply (qwen_stream), a long-chat compaction, and a "draft …" dictation still work.

🤖 Generated with Claude Code