fix: robust response parser + domain-aware personas

[arvarik]

Summary

Two critical blackboard architecture failures identified from post-mortem of task-caccf02b are fixed.

Problem

1. Empty refs fields — agents write **Refs**: [e-3, e-4] in body prose but leave the structured refs JSON field empty. The routing and convergence machinery depends entirely on this field (_deterministic_fallback, _infer_phase, salience scoring).

2. Bundled entries — planner wrote 4 plans in one entry; critic wrote 4 critiques in one entry. Breaks selective referencing and prevents the CU from routing critique authors back to their own work.

Root cause: persona prompts described fields without showing format examples. Parser trusted agent output completely — no extraction fallback.

Changes

core/response_parser.py (new)

587-line standalone parser with 7 extraction strategies applied in priority order:

Pattern	Fix
"refs": ["e-3"] in JSON field	Pass through (happy path)
**Refs**: [e-3, e-4] in body	Regex → refs field
refs=[e-3] in body	Regex → refs field
(refs: e-3) in body	Regex → refs field
Bare e-N mentions in body prefix	Heuristic → validated against known_ids
Multiple entries in one body	Split by entries array, --- delimiters, or ## Section headers
Decider JSON code-fence wrapping	Unwrap + merge inner refs
type: finding + rebuttal signals in body	Auto-promote to rebuttal
Confidence as string / missing	Parse + hedging-word heuristic
action: decline / action: clean	Pass through unchanged

core/variants/traditional.py

Replaced naive parse_agent_response with delegation to parse_entries. Added known_ids parameter.

core/orchestrator.py

Threads known_ids (set of board entry IDs from the already-fetched snapshot) into parse_agent_response to enable validated prose ref extraction.

models/personas.py (full rewrite)

All 5 constant role personas + expert generator rewritten with explicit JSON output contracts, worked examples, confidence calibration tables, and anti-pattern sections:

• Planner/critic: require entries array, one entry per idea
• Expert: clear rule — if refs contains critique ID → type: rebuttal
• Decider: body must be plain prose, never a JSON code block; refs must be exhaustive
• AG prompt: domain-specific experts only, generic roles (Domain Analyst, Systems Thinker) forbidden

core/config.py + core/blackboard.py

Node-name normalization: NODE_URL_TO_NAME reverse map; publish_log normalizes raw HTTP URLs to friendly node-1/2/3 identifiers.

tests/test_response_parser.py (new, 37 tests)

Every production failure pattern from the board audit has its own test case.

Test Results

ruff check src/ tests/    → All checks passed
mypy src/                 → Success: no issues found in 34 source files
pytest daemon/tests/      → 446 passed, 1 skipped
pytest agent/tests/       → 39 passed
eslint + tsc              → clean
vitest                    → 118 passed
npm run build             → ✓

Architecture Impact

Before	After
e-5.refs = [] (body says **Refs**: [e-3, e-4])	e-5.refs = ["e-3", "e-4"]
Planner → 1 entry with 4 sub-goals	Planner → 4 separate plan entries
Critic → 1 entry with 4 critiques	Critic → 4 separate critique entries, each with own refs
Rebuttals typed as finding	Auto-promoted to rebuttal on body signals
Decider body = JSON code block	Unwrapped to plain prose
All confidence = 0.5 (flat)	Calibrated per agent + hedging heuristic

Closes the routing blind-spot: _deterministic_fallback and _infer_phase now have accurate refs to work from.