fix(ai-improve): tolerate prose-wrapped JSON from the model

[ssevera1]

Reproduced 3/3 in a row on Data-Science-Toolkit while verifying the pipeline: the selection call succeeded (context built, API responded), but Haiku sometimes reasons in prose ("Looking at this codebase, I need to identify...") instead of emitting bare JSON, despite SELECT_PROMPT's existing "ONLY valid JSON" instruction. ask_json() only tried json.loads() on the whole trimmed response, so any surrounding prose made it discard an otherwise-usable answer and skip the run entirely — safely (no bad file written), but silently.

What changed

• Stronger prompt instructions in both SELECT_PROMPT and WRITE_PROMPT: explicitly forbid reasoning/commentary, spell out that a caller parses the response programmatically
• Fallback extraction in ask_json(): if the full response isn't valid JSON on its own, scan for the first balanced {...} object anywhere in it and parse that instead of giving up
• Print more of the raw response on a genuine parse failure (800 → 2000 chars) for easier debugging next time

Same file, byte-identical across all 11 repos before this change — opening the same fix everywhere rather than just on the repo that surfaced it.

[claude]

Checked the fallback JSON extractor, the prompt escaping, and the downstream call sites. The balanced-brace scan only runs after a plain json.loads failure, so the happy path is unchanged; its one weakness (brace counting ignores JSON string context) fails closed — a premature } yields an unterminated string that json.loads rejects, returning None exactly as before, so it can never hand back a silently truncated file_content. Extracting a stray object from prose is still gated downstream by safe_target(), the required-keys check, and the SHRINK_FLOOR guard, and the new prompt text correctly doubles {{/}} for the .format() templates. Scope is one file and matches the stated intent.