docs: clarify 12b thought markers

experiments/reposhell-prompt-propensity/README.md

@@ -52,7 +52,12 @@ prompt style, not the explicit upper-bound wording.
 Thinking-high follow-up:
 
 - `LOCALPAGER_AGENT_THINKING=high` recorded `thinkingLevel: high` in all six
-  saved sessions, but the transcripts contained zero `type: "thinking"` blocks.
+  saved sessions.
+- Pi/localpager-agent did not save 12B reasoning as structured
+  `type: "thinking"` blocks. It did save Gemma thought-channel markers such as
+  `<|channel>thought` / `<channel|>` as ordinary text content. This also
+  happened in the 12B thinking-off runs, so the correct reading is: no
+  structured thinking blocks, but 12B thought markers leaked into text.
 - `hard-source-gate` stayed clean: 3/3 targets used exactly one successful bash
   call and emitted final JSON.
 - The actual evidence remained shallow for `hard-source-gate`: all three calls
@@ -63,9 +68,20 @@ Thinking-high follow-up:
 - Tool-output quality across the thinking-high run: 2 substantive reads, 9
   shallow lookups, 27 failed calls.
 
-Conclusion: thinking high does not fix reposhell quality here. The better next
-change is to steer required source checks toward one simple `rg`, `git grep`,
-or `sed -n` lookup, and to discourage `ls` as sufficient evidence.
+12B thought-marker audit:
+
+| Run | Setting | Structured thinking blocks | Thought-marker text parts | Non-empty marker payloads |
+|---|---|---:|---:|---:|
+| `run-20260615T165426Z` | default prompt, thinking off | 0 | 34 | 2 |
+| `run-20260615T165912Z` | strict wrapper, thinking off | 0 | 12 | 2 |
+| `run-20260615T171536Z` | strict production prompt, thinking high | 0 | 38 | 1 |
+
+Conclusion: the earlier shorthand "zero thinking blocks" was too narrow.
+Thinking high did not fix reposhell quality here, and 12B's thought markers were
+mostly empty or formatting leakage rather than useful source-grounding
+reasoning. The better next change is to steer required source checks toward one
+simple `rg`, `git grep`, or `sed -n` lookup, and to discourage `ls` as
+sufficient evidence.
 
 The soft prompt-only changes did **not** increase reposhell use under these
 conditions. Harsher wording did increase tool use, but with quality problems:

experiments/reposhell-prompt-propensity/run-20260615T171536Z/README.md

@@ -12,6 +12,8 @@ Goal: test whether prompt wording can increase voluntary read-only repository in
 - Per-call timeout ms: 300000
 - Target count: 3
 - Variant count: 2
+- Thinking: `LOCALPAGER_AGENT_THINKING=high`; all six sessions recorded
+  `thinkingLevel: high`.
 - Reposhell was exposed through the Localpager agent runtime.
 - Localpager repo files were not edited; generated prompts, logs, sessions, and summaries are under this scratch experiment directory.
 
@@ -22,6 +24,12 @@ Goal: test whether prompt wording can increase voluntary read-only repository in
 | hard-source-gate | 3 | 3 (100.0%) | 3 (100.0%) | 3 | 3 | 0 | 12364 |
 | explicit-bash-upper-bound | 3 | 3 (100.0%) | 3 (100.0%) | 35 | 8 | 0 | 27057 |
 
+Thinking audit: Pi/localpager-agent recorded zero structured
+`type: "thinking"` blocks, but 12B emitted 38 Gemma thought-channel markers
+such as `<|channel>thought` / `<channel|>` as ordinary text content. One marker
+had non-empty payload. This means the run had thought-channel leakage in text,
+not structured thinking blocks.
+
 ## Variants
 
 ### hard-source-gate