COGNITIVE-CYCLE-3AGENTS-PILOT-V2

RETAKE of cognitive-cycle-3agents-pilot (v1, commit 78abbf9d6) with fixed parser + prompt examples. v1 diagnosis: strict [TYPE] content whitelist rejected most agent outputs (variants like [COUNT] x, OBSERVATION — x, OBS: x all marked SILENT), forcing 19/24 turns into single OBSERVATION bucket + 5 SILENT — the apparent «convergence» (P1 JS=0.000) was an implementation artifact, not a substrate property. 8/19 broadcasts even contained the literal word «content» (LLM copying template placeholder). v2 fixes: 1. Lenient parser: bracketed/colon/dash header patterns; normalize COUNT|NUMBER|ENUM→OBSERVATION, HYPOTHESIS|IDEA|GUESS→CLAIM, ASK|INQUIRY→QUESTION, SKEPTICAL|UNSURE→DOUBT, QUIET|NOTHING→SILENT. Unrecognized text → FREE_TEXT (new bucket, broadcasts content). Template-leak guard: «content» literal → FAILED_PARSE (separate diagnostic bucket, not silent). 2. Prompt examples: 5 concrete one-line examples for each type embedded in ACT specification; SILENT explicitly framed as valid choice not failure. 3. Smoke test: scripts/test_cc_parser.py exercises 21 representative inputs. Hypothesis tested: was v1's convergence finding (P1 JS=0.000, P4 cos=0.985) a substrate property (substrate genuinely flattens all asymmetry into one output style) OR a parser artifact (LLM produced diverse types, parser threw them away)? v2 with working parser distinguishes.