## Context

P3C-029 through P3C-033 require raw, reproducible processing, payload-rate,
fairness, cap, and impairment evidence. Existing focused tests cover the
framer, bounded queues, native Apollo fake, and fair pacer, but do not emit the
normative ten-minute or six-profile artifacts.

## Goals / Non-Goals

**Goals:**

- Run the three 20/50/80 Mbps encoded-payload profiles for ten wall-clock
  minutes each after a recorded warm-up.
- Measure the existing gateway framing path with a monotonic clock and retain
  compressed raw latency samples plus full summary statistics.
- Run the exact six Section 7.2 configurations once using a deterministic,
  bounded virtual packet discipline and retain configured and observed values.
- Exercise one real mTLS/QUIC fake-provider traversal for every media profile
  and reuse the real fair-pacer implementation for fairness and cap evidence.

**Non-Goals:**

- Live Apollo/macOS/firewall qualification, real encoder fidelity, codec
  processing, host network mutation, or multi-host scale.
- A production impairment framework, new gateway API, dependency, cgo, or
  sidecar.

## Decisions

- Implement the harness as an opt-in `go test` in package `gateway`. This
  keeps qualification access to the actual unexported fair pacer without
  adding a production API. Normal suites skip the long run unless an explicit
  absolute evidence directory is supplied.
- Use wall-clock duration and target-rate pacing for performance profiles.
  Measure only receive-to-framed-payload processing; pacing wait and raw-file
  writes stay outside the measured interval and are reported separately.
- Stream every raw sample into gzip-compressed CSV while retaining one bounded
  duration slice per profile for exact percentiles.
- Use a fixed-seed virtual FIFO for impairment. It records no host claim and
  identifies its queue discipline and deterministic topology explicitly.
- Treat any payload mutation, p95 above 5 ms, catalog mismatch, fairness error
  above 10%, cap excess above 5%, or step convergence beyond ten seconds as a
  hard command failure.

## Risks / Trade-offs

- [Local virtual impairment cannot prove deployed route behavior] → label every
  artifact deterministic and retain live Apollo/macOS/firewall as
  deferred-owner-e2e.
- [Raw samples can be large] → stream gzip output and bound in-memory samples
  to the exact profile packet budget.
- [Host load can invalidate latency] → record OS, architecture, Go version,
  timing overhead, actual duration, packet count, and observed bitrate; fail
  rather than substitute configured capacity for measured egress.