feat(perf): verify shared recipe bundles and document pinned CLI workflow

[plotarmordev]

Closes #25. Stacked on the observed-policy PR; this diff adds the shared bundle and workflow only.

Delivered

• Offline bundle verify MANIFEST --json, bounded closed schema with four workload entries and explicit DeepSeek/GLM decode/prefill mappings.
• Existing frozen sparkDash files stay unchanged. Materialized GLM variants may differ only in name and the explicit enable-thinking declaration; all other semantic differences are rejected even if hashes are recomputed.
• Raw and normalized workload identities, safe leaf paths/symlink checks and declared request/token ceilings. No runtime resolver, generation, network call or second collector.
• Shared pinned-source Linux installation instructions, including native ARM64 prerequisites, actual binary/build/lockfile identity and one canonical workload location. Concrete source pin a6729f9 is an existing ancestor rather than a self-referential hash.
• Thin existing CLI command workflows for both recipes, explicit precollection policies and manual reviewed report template. No shell orchestration, model discovery, restart/reconfiguration, cache flush, weight download, exporter or upload.

Verification

SIX and CYBERGLM reviewed the bundle independently; final integration also reviewed by SIX/HELP/CYBERGLM, including the documentation pin delta. Six new real-CLI bundle tests, plus exact-head fmt/Clippy/workspace tests/release build. Clean-source native Linux ARM64 build at the documented pin verified the manifest from another directory and exercised all four entries through baseline/candidate/repeat/decide: 528 synthetic loopback requests with deliberately permissive test policies, no inference calls. Integrated suite: 176 passed, 1 ignored.

Limits

These are source, CPU build and loopback protocol results, not DeepSeek or GLM live qualification. Their statuses remain separate; DeepSeek is first only after a coordinated serving window, GLM later. Existing benchmark eligibility is not PASS, and the new decision is scoped observed-policy compliance rather than statistical/capability certification. Public summaries require manual review; raw evidence and operator identifiers are not automatically safe to share. The documented pin contains the implementation; current installation instructions are in this later documentation commit.