XY-1155: Implement source-backed memory quality benchmark

[yvette-carlisle]

Summary

Implements the XY-1155 executable source-backed memory quality benchmark gate for ELF's final source-backed project memory program.

Changes:

• adds source_backed_quality JSON report surface and Markdown rendering
• adds validate-source-backed-quality so the benchmark command fails on non-pass source-backed quality gates
• adds cargo make source-backed-memory-quality JSON -> validate -> Markdown workflow
• adds a Context Pack activation/suppression fixture with structured routing decisions
• hard-fails required non-pass scenarios, required job trap/hard-fail usage, cross-scope leaks, journal-only authority claims, and incorrect/missing Context Pack routing decisions
• counts private_scope_leak as a scope violation for the source-backed quality gate
• documents benchmark evidence and review disposition

Decodex/manual intervention

Decodex was attempted first for XY-1155. Manual intervention was required after repeated Decodex app-server timeouts/no-effective-diff loops, including Timed out while waiting for app-server output at apps/decodex/src/agent/json_rpc/connection.rs:287 and retained attention after attempts 2/3. This PR is the manual completion path and should be rebound into Decodex review handoff.

Benchmark evidence

cargo make source-backed-memory-quality now runs:

1. JSON generation
2. validate-source-backed-quality
3. Markdown publication

Latest local artifact: tmp/source-backed-memory-quality/report.json:

• source_backed_quality.result_state = pass
• hard_fail_passed = true
• expected evidence recall: 1.0
• precision@5: 0.414
• irrelevant context ratio: 0.0
• source-ref coverage: 1.0
• stale suppression: 1.0
• correction persistence: 1.0
• delete/tombstone suppression: 1.0
• unsupported claim rate: 0.0
• cross-scope leak count: 0
• journal-only authority claim count: 0
• Context Pack activation precision/recall/trace coverage: 1.0 / 1.0 / 1.0
• Context Pack routing decisions: 7 total, 7 traced, 0 incorrect

The aggregate report still preserves typed non-pass competitor/runtime boundaries and does not support an unqualified product-runtime leaderboard claim.

Review

A first read-only skeptic review blocked the change on three P1s: missing executable gate, insufficient hard-fail coverage, and tag-derived Context Pack metrics. The implementation was updated, and a second read-only skeptic review returned pass with no P0/P1 findings.

Validation

• cargo make checks
• cargo make source-backed-memory-quality
• python3 scripts/check-docs.py
• git diff --check