Cut set-enrichment cost with structured capability matching

## Problem

Enriching one set costs ~$3.85: guide analysis ($0.42), per-card analysis ($0.43), and 1117 per-pair relationship-validation model calls (~$3.00) on HOB. Validation scales with the capability cross-product, not the card count, making per-set enrichment roughly 3x the promised $1-2 estimate.

## Goal

Enrichment drops to ~$1.4-1.6/set by extracting structured capability parameters once per card and resolving most candidate pairs locally, reserving batched model verification for the ambiguous residue.

## Scope

Card-capability extraction schema v2 (action/zone/qualifier fields), local deterministic pair matcher, residual-only batched validation, work-store identity migration.

## Out of scope

Guide-extraction changes; scoring-side consumption of typed prerequisites (#508 records stay compatible).

## Child issues

- [ ] Card capability schema v2 with structured mechanism parameters
- [ ] Local pair matcher with residual routing to batched validation

## Acceptance Criteria

- [ ] AC1: Full HOB enrichment re-run completes with <= 60 model calls total for validation
- [ ] AC2: Published HOB artifact passes the frozen benchmark gate (accepted/rejected/uncertain verdicts match benchmark.json within the agreed tolerance)
- [ ] AC3: Measured cost per set <= $1.60 at current token prices

## Epic classification

- **Overall size:** L
- **Overall orchestration risk:** High
- **Reason:** Cross-subsystem response-contract migration with a benchmark-gated quality bar and stored-work invalidation.

Start by reading the two child issues and the existing HOB enrichment flow, then inspect benchmark.json and the work-store identity handling. Done means the HOB rerun stays within 60 validation model calls, matches the frozen benchmark tolerance, and costs no more than $1.60 per set.

Written by the indexing model from the issue text.