Independent, abuse-resistant bandwidth measurement and enforcement (remainder of #30)

[flo2517]

Split off from #30 once its first slice (units + Agent capability reporting + scheduler reservation/overcommit prevention) landed. This issue tracks the remaining, substantially larger part of #30's acceptance criteria:

• Agent reports bounded signed usage summaries (not just a declared static capacity) -- an ongoing measurement + signing protocol, not yet designed.
• Independent validators run abuse-resistant multi-point tests rather than trusting the provider's self-reported ingress/egress figures -- the bandwidth capacity landed in Benchmark and enforce provider bandwidth capacity #30's first slice is entirely operator-declared (the same trust boundary as price/reputation self-declarations), with zero independent verification. This is the core 'benchmark' half of Benchmark and enforce provider bandwidth capacity #30's title that hasn't been built yet.
• WireGuard overhead (the lease-gated overlay from Add WireGuard overlay networking after lease authorization #38) and regional endpoint selection are not accounted for anywhere in bandwidth capacity or fit scoring.
• Workload-level rate limits are not enforced -- a workload can request bandwidth and have it reserved, but nothing currently throttles its actual usage against that reservation.
• Test coverage for congestion, asymmetric links, spoofed results, and network partitions -- none of this exists yet since there is no active measurement subsystem to test.

Likely needs its own ADR (a new validator active-measurement subsystem is a structural addition, not an incremental change) before implementation starts.

🤖 Filed by Claude Code during autonomous v0.3 milestone work, per user's standing instruction to open issues for scope discovered along the way.

[flo2517]

Folding this into #78: this issue's ask (independent validators running abuse-resistant multi-point tests, rather than trusting the operator-declared bandwidth figure from #30) is a subset of what #78/ADR-013's challenge-loop slice builds anyway — the same daemon (cmd/networkvalidator), with bandwidth as one more ScoreDimension::Network it measures, not a separate subsystem. Building these as two independent tracks risks two different architectures for the same job.

Keeping this issue open specifically for the parts that are genuinely bandwidth-specific and not covered by #78's general framing: congestion, asymmetric-link, spoofed-result, and partition test scenarios. Will close as subsumed once #78's challenge loop supports the Network dimension, if these scenarios are covered there too.

🤖 Claude Code

[flo2517]

Status update: ADR-015's core "independent bandwidth measurement" slice is merged (PR #90, ADR text).

What's now real:

• A new MeasureBandwidth gRPC RPC (protocol/proto/openinfra/agent/v1/agent.proto), distinct from SolveChallenge's existing TYPE_NETWORK.
• agent-api's handler: bounded payload (MAX_BANDWIDTH_PROBE_BYTES = 8 MiB), real per-caller rate limiting (10 calls/60s, keyed by the caller's mTLS-derived Ed25519 public key), domain-separated Ed25519 signing over the response.
• The Network Validator's challenge loop now calls MeasureBandwidth instead of the generic liveness/correctness check for the Network dimension, scoring pass/fail against the provider's own declared ResourceCapability.Bandwidth with a 70% tolerance.
• Independently re-verified before merging: byte-for-byte signing-construction match between the Rust and Go sides, real signature-tampering tests, real unreachable-agent tests, buf generate producing a zero diff, and blockchain/ confirmed untouched.

Leaves this issue open for its explicitly out-of-scope remainder, per ADR-015 §"Consequences":

• WireGuard overhead accounting (a lease-gated overlay adds its own throughput ceiling this raw Agent-to-validator measurement doesn't see).
• Regional endpoint selection.
• Workload-level rate limit enforcement (this measures capacity; it does not throttle a running workload's actual usage against its reservation).
• The adversarial test suite (congestion, asymmetric links, spoofed results, partitions).

One design note carried over from PR review, worth tracking here rather than losing it: a provider with no declared bandwidth in its heartbeat cache currently auto-passes the tolerance check (nothing to compare a measurement against) rather than getting an explicit "unscored" state. Narrow edge case today since heartbeats normally always populate this from real sysinfo data, but worth hardening alongside the adversarial test suite above.