neural-fabric: build the smoke/full harness + re-enable it in CI

[mdheller]

Context

PR #69 scoped neural-fabric-ci to neural-fabric-static because the smoke/full targets referenced a harness that was never committed (validate-neural-fabric-results.py, check-capacity-bounds.py, tests/neural_fabric/, run_suite.py, and the intervention-event schema). This builds that harness for real, against the actual reference primitives (may_wigner + hopfield), and re-enables the gate.

What's added

• scripts/check-capacity-bounds.py — May-Wigner capacity gate: s·√(m·C) → classify (ok/warn/error/stop), exit non-zero past --max-class.
• scripts/validate-neural-fabric-results.py — jsonschema conformance for families / experiments / results / events, plus the core doctrine invariant (weights_updated is always false) and result/event → experiment cross-references.
• research/activation-time-targeting/code/run_suite.py — deterministic, CPU-only suite:
  • NF-ATT-001 — Hopfield logit-boost steering (monotone target-probability lift, no weight update); emits intervention-event.v1 audit records.
  • NF-ATT-002 — May-Wigner capacity sweep locating the warn/error/stop boundaries.
• tests/neural_fabric/ — 25 known-answer tests for the primitives + invariants on the committed reference artifacts.
• schemas/neural-fabric/intervention-event.v1.json — restores the audit-event schema dropped in integ: land 2 clean backlog branches #69 (the doctrine governs every intervention by an audit event). (This resolves the "consider whether intervention-event was intended" question: yes — it's the audit record for each intervention.)
• Committed reference family / experiments / results / events.

mk / workflow

• neural-fabric-static re-adds the intervention-event schema check.
• neural-fabric-ci: neural-fabric-static neural-fabric-smoke (smoke re-enabled).
• neural-fabric-full regenerates the committed artifacts in place (deterministic) → doubles as a reproducibility check.
• Workflow installs jsonschema.

Verification

• make -f mk/neural-fabric.mk neural-fabric-ci → 8 artifacts valid, capacity PASS, 25 tests passed.
• Full repo pytest → 92 passed. ruff clean. No __pycache__ committed.