…ojections

Replace the two drifting category→metadata tables (EDGE_CATEGORY_METADATA
in schema/edges.ts and the drifted CATEGORY_POLICY) with a single
EDGE_CATEGORY_METADATA in policy/category-policy.ts. The old pair
disagreed on impact direction for proof/support; the consolidated table
carries endpoint roles, impact direction+strength (none/advisory/cascade),
criteria-help, and projection effect in one place.

Add two anchor-relative projections derived from it:
- projection/labels.ts: direction-aware semantic phrasing (two-tier).
- projection/direction.ts: upstream/downstream/lateral for the
  reconciliation flow, derived from the impact columns.

Wire node-neighborhood renderer to group incident edges by directional
relation and label each line semantically; retire the inline endpoint
role-token leak (graph-slice + edges.ts return to baseline). Reconcile
GRAPH_MODEL.md and graph/README.md to the new shape.

Amp-Thread-ID: https://ampcode.com/threads/T-019ea862-2101-73ce-8c08-09aca617f03a
Co-authored-by: Amp <amp@ampcode.com>

Add edge-spread/hub-neighborhood.json: one requirement anchor wired to
naturally-typed neighbors covering every edge category in both relation
directions, both proof/support stances, the three realization refinements
(expresses / implemented by / established by), hard vs soft impact, a
lateral association, and an ambient cross-neighbor edge.

Lock three renderer previews exercising the projection layer:
- neighborhood-hub-REQ1: integrated directional + semantic render (hops=1)
- neighborhood-hub-REQ1-hops2: ambient-edge counting at hops=2
- graph-slice-category-directions-full-debug: anchorless flat edge render
  across all categories (regression-guards the reverted slice renderer)

Amp-Thread-ID: https://ampcode.com/threads/T-019ea862-2101-73ce-8c08-09aca617f03a
Co-authored-by: Amp <amp@ampcode.com>

…shape-only renderer previews

Split verification along the design seam: labels and directionality are pure
functions, so they own correctness; renderers own only output shape.

Projection unit tests, now the full executable lookup matrices:
- labels.test.ts: every Tier-1 (category, anchorRole, stance) base cell +
  every Tier-2 (category, sourceKind, targetKind) realization refinement.
- direction.test.ts: every category's impact axis and the neighbor
  relation+strength from both anchor endpoints.

Renderer tests consolidated into one home, src/renderers/graph/previews.test.ts:
- the locked __previews__/*.md snapshot IS the assertion; positive
  'looks like X' string matches removed (the projection matrices prove those).
- only cross-cutting shape invariants kept: no raw structural vocabulary leaks
  into anchored projections, and large-graph slices stay bounded.
- adds neighborhood-hub-REQ1-bounded preview locking maxEdges omission.
- retires graph-slice.test.ts, node-neighborhood.test.ts,
  node-neighborhood-spread.test.ts, and the .pi wrapper node.test.ts (the
  wrapper is a pure pass-through; maxNeighbors dead option dropped).

Also lands the architecture-boundary fix from the parallel agent: the real
fixture-reads.test-support implementation moves to src/graph/, leaving a
re-export shim under src/renderers/graph/.

All src/graph + src/renderers + architecture tests pass (197). Unrelated
pre-existing failure in src/probes/structured-exchange-ordering-proof remains
(temp extension load), untouched.

Amp-Thread-ID: https://ampcode.com/threads/T-019ea862-2101-73ce-8c08-09aca617f03a
Co-authored-by: Amp <amp@ampcode.com>

…PLAN.md

First fixture hand-derived from real planning prose rather than synthetic
placeholders, proving the full loop: prose -> graph fixture -> the real
propose-graph validator (seedFixture -> CommandExecutor) -> renderers.

.fixtures/seeds/brunch-self/spec-graph.json: one planning_ready spec describing
Brunch itself across all four planes. As a by-product of being faithful it
covers every node kind and every edge category (both proof/support stances) and
includes a supersession lineage. The grouped-list render shows 54 nodes / 30
edges from 55 / 31 authored because the active-context projection (D60-L)
correctly hides the superseded predecessor requirement and its dangling edge.

src/renderers/graph/previews.test.ts: locks a whole-spec grouped-list plus
REQ1 (one-authority requirement) and MOD1 (CommandExecutor, hops=2)
neighborhood previews. Seeding through readGraphSliceFixture also gates
structural legality — the test fails if any node/edge is illegal.

Validated: src/graph + src/renderers (195 tests) pass.
Amp-Thread-ID: https://ampcode.com/threads/T-019ea862-2101-73ce-8c08-09aca617f03a
Co-authored-by: Amp <amp@ampcode.com>

…e) + validate-fixture CLI

Hand-derive structurally-legal spec graphs from four external projects'
prose docs to exercise the propose-graph validator and graph renderers
beyond the brunch-self example:

- dumpchat   41 nodes / 33 edges  (docs/SPEC.md, mostly faithful intent/design)
- yamlbase   69 nodes / 65 edges  (memory/SPEC.md + PLAN.md)
- rd-loop    54 nodes / 35 edges  (README + concept docs, more projection)
- fable      67 nodes / 37 edges  (memory/SPEC.md + ROADMAP.md)

Each covers all four planes, every node kind, all eight edge categories,
both proof/support stances, and one acyclic supersession lineage.

Add src/graph/validate-fixture.ts, a standalone CLI that seeds one fixture
through the real CommandExecutor and reports authored/stored/active-context
counts, so fixtures can be authored and validated without test-file contention.

Amp-Thread-ID: https://ampcode.com/threads/T-019ea862-2101-73ce-8c08-09aca617f03a
Co-authored-by: Amp <amp@ampcode.com>