fix: walk back round-1/2 conclusions — the cause was think=false all along

Seth asked "was this with think=false?" Yes — and that was the only
question that mattered. Everything I concluded in round 1 and round 2
was wrong.

Actual cause, isolated in round 3:
- At identical message state, gemma4:26b with think=false returns
  eval=4 (silent stop); with think unset or think=true, returns
  eval=165 and emits the correct tool call.
- Original round-1 write_file harness + think unset: 26B passes in
  8 iters, 20s. No mitigations needed.
- 31B dense and qwen3-coder:30b tolerate think=false; 26B MoE does not.

Red herrings (kept on-record in the bakeoff doc, not silently erased):
- Round 1: "write_file tool-call argument size" — wrong
- Round 2a: refuted the arg-size theory but for the wrong reason
  (still failed because think=false was still set)
- Round 2b: "cumulative tool-response context size" — truncating
  did make 26B pass, but by coincidence. Shorter context at the
  decision turn dodged the think=false side effect.

Why the existing "always think:false" guidance was misleading:
it was derived from AI_Visualizer (single-turn JSON pipelines) where
thinking tokens do eat num_predict invisibly. In multi-turn
tool-calling agents the channels are separate and the flag has a
different effect — catastrophic on 26B specifically.

Doc updates:
- GOTCHAS: replaced the 26B entry with the actual cause; scoped the
  original "Thinking Mode Eats Context" entry to single-turn pipelines
- SYNTHESIS: split the "Mandatory Ollama Settings" block into
  single-turn vs multi-turn variants; updated anti-patterns and
  quick-start checklist
- CORPUS_cli_coding_agent.md: revised pointer and config template
- docs/reference/bakeoff-2026-04-18.md: added Round 3 section with
  the correction notice at the top of the file and full diagnostic
  methodology

New artifacts: harness_no_think_flag.py, harness_write_no_think.py,
and 4 new log files demonstrating all three models pass when think
is left at default.

CORPUS_cli_coding_agent.md

@@ -6,16 +6,17 @@
 > `IMPLEMENTATIONS.md` chat-agent patterns (Simon) and pipeline patterns
 > (AI_Visualizer).
 
-> **Empirical follow-up:** `docs/reference/bakeoff-2026-04-18.md` — 2 rounds of
+> **Empirical follow-up:** `docs/reference/bakeoff-2026-04-18.md` — 3 rounds of
 > runs against a custom minimal CLI-agent harness on a fix-the-median-bug task.
-> **Round 1:** 31B clean (8 iters), Qwen3-Coder correct but chatty (15 iters),
-> 26B silently quits with zero edits. **Round 2 (diagnostic):** the 26B failure
-> is NOT about edit-tool-argument size — it's about **cumulative tool-response
-> context shape**. Capping tool responses ≤1200 chars makes 26B pass cleanly
-> *and* in the fastest wall time of any run (8.4s). Most production CLI agents
-> already truncate tool responses, so the issue may be invisible in them.
-> Read when: scoping which model to point an agent at, hitting an unexpected
-> tool-call halt, or writing a custom harness targeting the 26B MoE.
+> **Bottom line after the full investigation:** all three models (gemma4:26b,
+> gemma4:31b-it-q4_K_M, qwen3-coder:30b) pass the task cleanly in 8-14
+> iterations. The only real gotcha is **`think: false` in the Ollama payload
+> silently breaks 26B** in multi-turn tool-calling loops — contradicts the
+> older "always think:false" guidance which was derived from single-turn
+> pipelines. Round 1 and Round 2 pursued wrong hypotheses (edit-tool size,
+> tool-response size) before Round 3 isolated the actual cause. Read when:
+> configuring an agent payload, debugging a silent halt, or fact-checking
+> older think-flag guidance.
 
 ## TL;DR
 
@@ -90,7 +91,6 @@ The baseline settings from `SYNTHESIS.md` still apply. CLI coding agent-specific
 ```json
 {
   "model": "gemma4:26b",
-  "think": false,
   "keep_alive": "4h",
   "options": {
     "num_ctx": 32768,
@@ -103,7 +103,7 @@ The baseline settings from `SYNTHESIS.md` still apply. CLI coding agent-specific
 - `num_ctx: 32768` is the working minimum for repo-scale work. Agents interleave file reads, bash output, and edits; 4K will truncate the second `read_file`.
 - `num_predict: 4096` — single edits are short but the agent may emit a bash invocation + reasoning + tool call in one turn.
 - `temperature: 0.3` — per `SYNTHESIS.md` temperature table, "structured extraction" tier. Coding edits want low variance.
-- `think: false` — critical. `GOTCHAS.md` documents that Ollama 0.20+ thinking silently eats `num_predict` and drops tool calls. If an agent somehow injects `think: true`, you'll see empty responses.
+- **Do NOT set `think: false`.** Leave it unset (Ollama default). Verified 2026-04-18 that setting `think: false` silent-stops `gemma4:26b` at tool-decision turns in multi-turn loops. The older single-turn-pipeline guidance ("always think: false") does not apply here — see `GOTCHAS.md` § "`think: false` Kills Gemma 4 26B in Multi-Turn Tool-Calling Loops".
 - `keep_alive: 4h` — agent sessions have think pauses; avoid reload penalty.
 
 ### Streaming