V3 W2–W6 continuation + comma quorum measured + 5+3 council + StyleFamily dedup (D-TSC-1)

.claude/agents/5plus3-council.md

@@ -0,0 +1,161 @@
+# 5+3 Council — spec-first hardening for delicate / ambiguous work
+
+> **Operator directive (2026-07-10):** *"for delicate/ambiguous create the
+> specs so detailed that a 5+3 council doesn't divert the path; cast the 5
+> first, consolidate first, then run the 3, then fix, then consolidate —
+> otherwise the 5+3 becomes mushy."*
+>
+> Prior art: OGAR `.claude/agents/README.md` (the original 5 savants + 3
+> brutal reviewers) and this repo's `epiphany-brainstorm-council.md` (panel
+> selection). This card is the **sequencing-hardened** synthesis of both —
+> the anti-mush protocol is the contribution.
+
+## When to convene (and when NOT to)
+
+Convene the council for **delicate or ambiguous** work only:
+
+- a dedup/refactor that touches ≥3 crates or retires a public taxonomy
+- a spec whose wrong resolution silently corrupts downstream sessions
+  (canon entries, LE-layout adjacent decisions, classid/mask semantics)
+- promoting a CONJECTURE toward canon when the probe alone can't close it
+
+Do NOT convene for: grindwork (one Sonnet worker + guardrails §1),
+mechanical probes with pre-registered gates (just run them), anything a
+single existing specialist card already owns end-to-end, or anything an
+operator ruling has already decided (rulings are frozen; the council
+verifies compliance, it never re-litigates).
+
+## The iron sequencing (anti-mush)
+
+```
+Phase 0  ORCHESTRATOR writes SPEC v1        ← the real work happens HERE
+Phase 1  cast the 5 savants, PARALLEL       ← they verify/harden, never design
+Phase 2  CONSOLIDATE FIRST → draft v2       ← orchestrator only; before ANY review
+Phase 3  cast the 3 reviewers, PARALLEL     ← they see ONLY draft v2, never raw fan-out
+Phase 4  FIX                                 ← apply verdicts; BLOCK = stop or re-spec
+Phase 5  CONSOLIDATE → ratified v3 + commit ← board hygiene same-commit
+```
+
+**Why the order is load-bearing (the mush failure modes):**
+
+| Skipped step | What goes mushy |
+|---|---|
+| Spec too thin (Phase 0 rushed) | 8 agents each invent a different architecture; consolidation becomes design-by-averaging |
+| Reviewers see raw savant output | 3 reviews of 5 divergent drafts = 15 cross-products of opinion; nothing is attackable |
+| 5 and 3 cast together | reviewers attack a moving target; savant findings and review verdicts interleave into soup |
+| Fix before consolidate | point-fixes to individual findings reintroduce the conflicts consolidation exists to resolve |
+| No final consolidate | the ratified artifact is a thread of patches, not a spec; the next session inherits the mush |
+
+## Phase 0 — the SPEC (the council-can't-divert bar)
+
+The spec is written by the orchestrator (main thread, full depth) BEFORE
+any agent is cast. A spec meets the bar when every section below is
+present and a savant could answer its questions without making a single
+design decision:
+
+1. **FROZEN DECISIONS** — numbered list. Operator rulings, iron rules,
+   prior council verdicts that apply. Each cites its source
+   (EPIPHANIES id / CLAUDE.md section / plan row). *The council may flag
+   a frozen decision only as VIOLATES with file:line evidence — never
+   re-open it on taste.*
+2. **INPUT INVENTORY** — every file:line the work touches, with shape
+   (type/variant counts, consumers, conversions). No "somewhere in
+   planner/" — exact paths. If the inventory needs discovery, that
+   discovery is a PRE-spec Explore task, not council work.
+3. **THE PROPOSED RESOLUTION** — the design, fully committed: target
+   shapes, migration steps in order, what is deleted / kept / aliased,
+   feature-gate and version-gate treatment (I-LEGACY-API-FEATURE-GATED).
+4. **NON-GOALS** — explicitly out of scope, each with one line of why.
+5. **PRE-REGISTERED GATES** — pass/fail criteria decided before any
+   agent runs (test counts, clippy -D warnings, parity assertions,
+   field-isolation matrix where layout is touched).
+6. **PER-SAVANT QUESTION SETS** — 3–6 numbered questions per lens,
+   answerable YES/NO/VIOLATES-with-evidence. This is what makes Sonnet
+   sufficient for the savants: bounded input, fixed output shape.
+
+## Phase 1 — the 5 (research savants, parallel, single lens each)
+
+Default panel (swap lenses per domain; declare swaps in the spec header):
+
+| # | lens | default card / charter | answers |
+|---|---|---|---|
+| 1 | prior art | `prior-art-savant` | is any part already shipped/named elsewhere? duplicate E-ids? |
+| 2 | iron rules | `iron-rule-savant` | YIELDS/VIOLATES per iron rule + AP1-AP9 |
+| 3 | code truth | runtime-archaeologist charter (via general-purpose) | is every file:line claim in the spec REAL? CODED vs CLAIMED vs ABSENT |
+| 4 | cascade impact | `cascade-impact-savant` | every file/test/doc/board row that must change; mandatory vs follow-up |
+| 5 | different views | `creative-explorer-savant` | the strongest alternative reading + the second-order consequence — WITHOUT redesigning |
+
+Casting rules:
+- **Model: Sonnet by default.** The detailed spec is precisely what makes
+  each lens grindwork-shaped (one bounded input, one fixed output shape).
+  Escalate an individual savant to Opus ONLY when its lens is genuinely
+  multi-source accumulation (e.g. prior-art across ~100 board/knowledge
+  docs). Never haiku (workspace floor).
+- Each savant receives: the FULL spec + its question set + the output
+  contract below. It does NOT receive the other savants' briefs.
+- **Output contract (per savant):** ≤10 findings, each =
+  `(question #, verdict from fixed vocab, file:line evidence, ≤2 sentences)`.
+  Verdict vocab: `CONFIRMS / VIOLATES / GAP / PRIOR-ART-AT / RISK`.
+  No prose essays. No redesigns. A savant that wants to redesign files a
+  single `RISK` finding naming the concern and stops.
+- Savants are read-only. The orchestrator is the only writer.
+
+## Phase 2 — consolidate FIRST (orchestrator, main thread)
+
+Before any reviewer exists: merge the ≤50 findings into **draft v2** =
+spec v1 + a change ledger:
+
+- every `VIOLATES` → spec amended or the frozen decision escalated to
+  the operator (never silently overridden)
+- every `GAP` → filled with a decision (committed, not optional)
+- every `PRIOR-ART-AT` → reuse wired in, duplication removed
+- conflicting findings → resolved by the orchestrator with one line of
+  why; the losing finding is recorded, not deleted (anti-collapse)
+- `RISK` findings → either absorbed as a gate or explicitly accepted
+
+Draft v2 is a single self-contained document. Raw savant output is
+banked (scratchpad or AGENT_LOG line) but never forwarded.
+
+## Phase 3 — the 3 (brutal reviewers, parallel, on draft v2 ONLY)
+
+| # | reviewer | charter (from OGAR; run via general-purpose if no local card) | blocks on |
+|---|---|---|---|
+| 1 | `overclaim-auditor` | every claim's grade matches its evidence; absolute words (`guarantee`/`proven`/`cannot`) only on [G] | grade inflation |
+| 2 | `dilution-collapse-sentinel` | no two motifs conflated (dilution); no valid leg deleted because a facet was wrong (collapse) | either |
+| 3 | `firewall-warden` | non-negotiables: no German PII labels, no model identifier in artifacts, no hot-path serialization, no prohibited shell, board hygiene same-commit | any hit |
+
+Casting rules:
+- **Model: Sonnet default** (draft v2 is self-contained); Opus when the
+  review requires independent codebase verification beyond the draft.
+- Output contract: per spec section, verdict `PASS / FIX(P1|P2) / BLOCK(P0)`
+  + evidence. "Looks good" without naming why each section earns PASS is
+  a malformed review — recast.
+- Reviewers never see Phase-1 raw output and never talk to each other.
+
+## Phase 4 — FIX (orchestrator)
+
+- Every `BLOCK` → resolve or return to Phase 0 (re-spec); never argue a
+  BLOCK away in the commit message.
+- Every `FIX` → applied, with the change ledger extended.
+- If two reviewers conflict, the stricter verdict wins by default;
+  overriding the stricter one requires an operator escalation line.
+
+## Phase 5 — final consolidate + commit
+
+Ratified **v3** = the executable spec. Then: implement (or hand to
+workers with guardrails §1), run the pre-registered gates, and land with
+board hygiene in the SAME commit (EPIPHANIES if a finding emerged,
+STATUS_BOARD row flip, AGENT_LOG entry naming the council run: which 5,
+which 3, verdict counts, what changed between v1→v2→v3).
+
+## Token-economy summary
+
+| role | model | why |
+|---|---|---|
+| Phase 0 spec + Phase 2/4/5 consolidation | main thread (full depth) | accumulation by definition |
+| the 5 savants | Sonnet (Opus only for true multi-source lenses) | the detailed spec bounds them into grindwork shape |
+| the 3 reviewers | Sonnet on the self-contained draft; Opus if codebase re-verification needed | one input, fixed output |
+| pre-spec discovery (inventory) | Explore/Sonnet | pattern matching |
+
+If the whole item is small enough that writing the spec costs more than
+doing the work twice — it was not council-grade; do it directly.