Document canonical Eval Lab failure and workload boundaries

[ashishki]

Outcome

Reconciles gdev-agent documentation with the published Eval Ground Truth Lab result for exact revision 0e4c5f0fd50382bbf12ffd35cfca4632384fb0cc.

The local 55-case conformance suite passes, while the unchanged public 100-case release gate correctly fails. The evidence pack is content-addressed as sha256:656face21f27b496d4d3e8bb0b588824f5737d122c1275c710f3e5b15ff94b4b and is published under Eval Lab v0.2.0.

Changes

• link the exact canonical result, sealed terminal record, manifest, dataset card, benchmark protocol, and human-review protocol
• replace stale “challenge not executed” text with the measured FAIL outcome
• define current maturity, portfolio ownership, product boundary, and non-goals
• remove technology badges and pilot/production-adjacent framing
• reconcile 55-case local conformance, 100-case release gate, and 180-case internal regression scopes
• add documentation assertions against framing regressions

Verification

• CI-scope Ruff check: pass
• CI-scope Ruff format: 97 files
• tests: 310 passed
• deterministic internal eval gate: 180/180, pass (run before the final docs-only boundary commit; executable code and cases are unchanged)
• documentation link and boundary assertions: pass

No production, pilot, external-user, live-traffic, or customer claim is made.