[P1][Post-0.5][Qualification] Record comparative performance, application regression, and release evidence

[mightnent]

Parent qualification epic: #79

Outcome

Produce comparable colocated and remote workspace performance distributions, run the complete application/agent regression matrix, and package secret-free release evidence tied to exact code and image artifacts.

Current baseline

The repository already provides workspace benchmark metadata, Kasm browser first-frame/input-to-paint collection, and deterministic Office regression tooling. It does not yet record the complete colocated/remote distributions and application matrix required here.

Execution

• Priority: P1 post-0.5 release evidence.
• Deployment profiles: Colocated customer-managed/worktree and representative hosted remote required.
• Worktree boundary: Benchmark orchestration, application/agent qualification matrix, evidence schema, secret scanning, and reports. Avoid performance-driven removal of required product capabilities.

Scope

• Record startup, desktop-ready, first-interactive-frame, input-to-paint, rendered-frame, bandwidth, CPU, and memory distributions separately for colocated and remote routes.
• Fix the agreed CPU/memory, browser, network, workspace profile, selected agents/applications, and persistent-home fixtures for meaningful comparison.
• Run DOCX/XLSX/PPTX roundtrip and reviewed GUI fidelity, Hermes Desktop/CLI, Claude Desktop/Cowork where supported, persistence, clipboard, agents, connectors, governed egress, and disabled-KVM behavior.
• Record unavailable measurements explicitly rather than as zero or inferred success.
• Produce a machine-readable manifest containing exact Git SHA, immutable image digests, topology, allocation, fixture versions, sample counts, commands, outcomes, and evidence hashes.
• Scan manifests, logs, traces, screenshots, and reports for launch URLs, grants, tokens, certificates, credentials, private targets, and customer data.
• Add Full C two-node latency/recovery evidence as [P1][Post-0.5][Epic] Productionize the Full C multi-node workspace runtime #82 scenarios become available.

Non-goals

• Removing Office, agents, Cowork, policy, connectors, or isolation controls to improve benchmark results.
• Treating local split-node latency as representative hosted production performance.
• Claiming GUI fidelity solely from structural document conversion.

Definition of success

• Colocated and representative remote distributions are recorded separately with agreed fixtures and adequate sample counts.
• The complete Office/Hermes/Claude/persistence/clipboard/agent/connector/egress/disabled-KVM matrix has explicit pass, fail, or unavailable outcomes.
• Every result is tied to the exact Git SHA, immutable image digest, topology, allocation, and evidence hash.
• Secret/private-target scanning passes and no customer data is used.
• Focused benchmark/regression suites, applicable browser qualification, and npm run verify:quick pass on the exact evidence SHA.