docs(openspec): archive gateway audit remediation

2026-07-30 07:39:18 +07:00
parent 9d627413a0
commit f292bd3599
14 changed files with 97 additions and 39 deletions
@@ -0,0 +1,2 @@
+schema: spec-driven
+created: 2026-07-29
@@ -0,0 +1,50 @@
+## Context
+
+The implementation already contains a source-shaped Apollo fake, native recovery, bounded gateway queues, fair pacing, Verse framing/QUIC, independent client support, lifecycle reporters, and low-cardinality telemetry. Audit defects arise where those existing pieces are bypassed or not connected.
+
+## Goals / Non-Goals
+
+**Goals:**
+
+- Reuse the existing production path for policy, lifecycle, recovery, telemetry, and qualification.
+- Delete duplicate qualification simulation.
+- Preserve all trust, cleanup, and resource bounds.
+
+**Non-Goals:**
+
+- Add codecs, provider transports, dependencies, or a generic lifecycle/telemetry framework.
+- Claim live Apollo/macOS/firewall interoperability.
+- Run the normative qualification before immutable consumer resolution.
+
+## Decisions
+
+- Format ANNOUNCE from `ProviderStreamPolicy` using the pinned Moonlight common-c bitrate and codec attributes. H.264 and HEVC with audio enabled are supported; AV1 and audio disabled fail before management/network readiness.
+- Advertise ordered registered `hevc-opus` and `h264-opus` profiles and delegate policy-compatible selection to generated Protocol intersection behavior. There is no Data Plane capability grammar.
+- Before `/applist` or `/launch`, validate the selected policy against source-backed Apollo `/serverinfo` codec flags and HEVC luma bounds plus reviewed source limits for dimensions, frame rate, bitrate, and audio where Apollo exposes no dynamic field. Reject rather than cap or downgrade.
+- Quiesce provider media sockets and the bounded forwarding path before emitting an existing terminal or disconnected event. Reuse current cleanup/release/reporting machinery and its cleanup-pending result; final control delivery has no fixed drain delay.
+- On a full audio FEC map, evict the oldest block according to existing block ordering and increment existing drop telemetry.
+- Carry provider receipt and queue-enqueue timestamps through the existing bounded media value. Queue residence, active processing, and scheduler pacing are sampled separately, once per complete provider media unit.
+- Sample existing process counters at heartbeat time; calculate rate from byte and monotonic-time deltas while leaving configured capacity in registration.
+- Build qualification on source-shaped pinned-mTLS Apollo management, encrypted RTSP, ENet, and provider UDP plus the public QUIC client path. Per-traversal stage deltas replace the standalone codec/parser and arithmetic impairment simulator; short smoke gates freeze the wiring, while normative durations remain deferred.
+- Preserve the production fair-pacer schedule across short host-timer overshoots so
+  measured allocation can catch up within the already bounded provider queue
+  instead of accumulating timer granularity as lost capacity.
+
+## Risks / Trade-offs
+
+- [Apollo cannot represent disabled audio truthfully] → Reject it rather than silently streaming stereo.
+- [Provider event races with queued or new media] → Quiesce ingestion and serialize forwarding with terminal event delivery before cleanup.
+- [Counter reset or zero elapsed time] → Emit zero measured rate and establish a new baseline.
+- [Corrected qualification is more expensive] → Run only short smoke tests until the immutable candidate is frozen.
+- [Apollo exposes incomplete dynamic capability detail] → Use only source-backed fields and explicit reviewed bounds; never infer support by silent capping.
+- [Pacer catch-up can emit a short burst after timer overshoot] → Clamp schedule
+  debt to five milliseconds in addition to the existing 16-packet provider
+  queue.
+
+## Migration Plan
+
+Land focused red/green repairs locally, verify through the temporary Protocol workspace, preserve old artifacts as superseded, and stop at the publication boundary. After a separately authorized immutable Protocol release is pinned, freeze inputs and run the corrected normative qualification once.
+
+## Open Questions
+
+None.
@@ -0,0 +1,28 @@
+## Why
+
+Fresh audit evidence shows the gateway ignores the immutable launch policy, leaves tunnels alive after provider termination/disconnect, can permanently stall audio after sustained loss, reports configured capacity as measured egress, and qualifies a standalone simulator instead of the production path.
+
+## What Changes
+
+- Negotiate registered decode profiles through the shared Protocol intersection, apply the effective policy to Apollo ANNOUNCE, and reject client or provider/source mismatch before launch (P3C-009, P3C-016, P3C-038).
+- Quiesce media immediately on provider termination/disconnect, deliver the final typed event reliably, and preserve bounded tunnel and durable lifecycle transitions including cleanup-pending (P3C-018–021, P3C-027).
+- Evict bounded stale audio FEC blocks so newer recoverable media continues (P3C-001, P3C-026).
+- Derive heartbeat egress and semantically separated queue, processing, and pacing observations from the production path (P3C-022, P3C-028).
+- Replace standalone processing/impairment simulation with a driver around the source-shaped provider, production queues/pacer/framing, QUIC, and an independent client (P3C-029–033, VER-008, VER-010).
+
+## Capabilities
+
+### New Capabilities
+
+- `apollo-stream-policy`: Native Apollo launch consumes the authenticated effective stream policy without downgrade.
+- `provider-session-lifecycle`: Provider terminal events close forwarding and report the correct durable lifecycle outcome.
+- `audio-fec-resilience`: Bounded Apollo audio recovery continues after permanently incomplete blocks.
+- `gateway-heartbeat-telemetry`: Authenticated heartbeat telemetry reports observed traffic and provider-path measurements.
+
+### Modified Capabilities
+
+- `gateway-qualification`: Normative evidence must traverse the production gateway path and retain raw resource, impairment, fairness, cap, and convergence observations.
+
+## Impact
+
+The native Apollo adapter, transport lifecycle, audio FEC state, heartbeat sampling, qualification driver, focused fixtures, and canonical qualification spec change. No dependency, cgo, sidecar, codec operation, direct provider route, or live interoperability claim is added.
@@ -0,0 +1,19 @@
+## ADDED Requirements
+
+### Requirement: Apollo launch consumes the effective policy
+The native Apollo backend SHALL derive ANNOUNCE resolution, frame rate, supported codec, selected bitrate, and audio profile from authenticated `ProviderSessionWork`, and MUST NOT substitute local defaults.
+
+#### Scenario: Supported HEVC policy reaches Apollo
+- **WHEN** provider work selects HEVC at 2560×1440, 120 FPS, 40000 Kbps, with audio enabled
+- **THEN** the encrypted ANNOUNCE carries those settings and the source-backed HEVC and bitrate attributes
+
+### Requirement: Provider policy cannot downgrade
+The gateway MUST use the generated Protocol intersection to select only a registered profile compatible with the immutable policy. It MUST reject invalid, unsupported, no-overlap, downgrade, audio-disabled, AV1, or provider/source-mismatched Apollo policy before `/applist`, `/launch`, or provider readiness because the current native path cannot truthfully honor those combinations.
+
+#### Scenario: Unsupported policy fails closed
+- **WHEN** authenticated provider work selects audio disabled, AV1, a codec outside the registered peer intersection, or a resolution, frame rate, bitrate, audio, or codec combination outside source-backed Apollo support
+- **THEN** setup fails before application discovery or launch without falling back to H.264, stereo, a cap, or another local default
+
+#### Scenario: Independent peers negotiate one registered profile
+- **WHEN** a production gateway and independent client advertise overlapping registered H.264 or HEVC profiles
+- **THEN** admission selects the first policy-compatible common profile using shared Protocol behavior
@@ -0,0 +1,8 @@
+## ADDED Requirements
+
+### Requirement: Bounded audio FEC state advances after loss
+The Apollo audio recovery window SHALL remain bounded and SHALL evict the oldest incomplete block when accepting a newer block would otherwise be rejected.
+
+#### Scenario: Newer complete block follows sustained loss
+- **WHEN** more than the bounded number of permanently incomplete audio blocks arrive before a complete newer block
+- **THEN** the oldest stale state is dropped, drop telemetry advances, and the newer encoded payload is relayed unchanged
@@ -0,0 +1,22 @@
+## ADDED Requirements
+
+### Requirement: Heartbeat egress is observed
+Authenticated gateway heartbeat telemetry SHALL calculate egress from monotonic transmitted-byte deltas over monotonic elapsed time and MUST NOT report configured capacity as measured traffic.
+
+#### Scenario: Controlled byte delta is sampled
+- **WHEN** transmitted bytes increase by a known amount during a known interval
+- **THEN** heartbeat egress equals the measured rate while configured capacity remains a separate registration value
+
+### Requirement: Required telemetry remains bounded and low cardinality
+The established authenticated path SHALL expose observed bytes, packets, drops, RTT, loss, jitter, queue delay, processing delay, pacing, reconnect, and provider state without session, route, credential, or payload labels.
+
+#### Scenario: Telemetry snapshot is published
+- **WHEN** the gateway emits a heartbeat after forwarding traffic
+- **THEN** it carries the bounded process-level observations and no high-cardinality or secret-bearing value
+
+### Requirement: Production delays have disjoint sample semantics
+Queue delay SHALL measure provider-queue residence, processing delay SHALL measure provider recovery plus framing and QUIC handoff work excluding queue and pacing, and pacing delay SHALL measure scheduler waiting only. The gateway SHALL advance processing samples once per complete provider media unit even when it emits multiple Verse frames.
+
+#### Scenario: Known production waits
+- **WHEN** one provider media unit has controlled enqueue, processing, and pacing intervals and fragments across multiple frames
+- **THEN** each cumulative total reports only its intended interval and exactly one processing sample is retained through authenticated Server persistence
@@ -0,0 +1,37 @@
+## MODIFIED Requirements
+
+### Requirement: Fixed media processing qualification
+The qualification harness SHALL drive pinned-mTLS Apollo management, encrypted RTSP, ENet, and provider UDP through native source validation, `readUDPMedia`, recovery/FEC, bounded production queues, the production fair pacer, Verse framing/QUIC, and a public or independent client decoder for 1080p60 H.264 at 20 Mbps, 1440p120 HEVC at 50 Mbps, and 4K60 HEVC at 80 Mbps. After a recorded warm-up, the frozen candidate SHALL run each profile for ten wall-clock minutes, preserve encoded payload bytes, retain every monotonic processing sample plus bounded CPU, memory, goroutine, allocation, and provider-queue observations, and report count, min, median, p90, p95, p99, max, mean, standard deviation, timing overhead, and observed bitrate. Processing begins at complete provider-unit receipt and ends at QUIC handoff, excluding client transit. Any bypass, payload mutation, wall-duration violation, bitrate outside both lower and upper bounds, or p95 above 5 ms SHALL fail.
+
+#### Scenario: Healthy fixed profile
+- **WHEN** a frozen candidate runs one fixed profile for the normative duration
+- **THEN** the harness emits compressed raw path and resource samples plus a summary tied to the exact command, topology, source commit, immutable Protocol version, environment, and payload hash
+
+#### Scenario: Processing gate failure
+- **WHEN** any production path stage lacks a per-traversal observation, payload integrity fails, duration or bitrate bounds fail, or measured p95 exceeds 5 ms
+- **THEN** the qualification command exits unsuccessfully without recording a passing candidate
+
+### Requirement: Bounded impairment qualification
+The harness SHALL run exactly the baseline, latency, jitter, loss, reorder, and constrained Section 7.2 profiles once by applying impairment at the source-shaped provider UDP boundary while traffic traverses the production gateway path. Baseline SHALL cover all three media profiles and the other profiles SHALL cover 1080p60. Each artifact SHALL retain raw impairment and queue observations and record tool version, exact command/configuration, environment, candidate commit, immutable Protocol version, direction, queue discipline, topology, fixed seed, and observed RTT, jitter, loss, reorder, throughput, drops, and capacity-step statistics.
+
+#### Scenario: Complete six-profile run
+- **WHEN** the frozen candidate runs impairment qualification
+- **THEN** one result exists for each named profile, with no Cartesian expansion and with raw observed rather than configured statistics from the real traversal
+
+#### Scenario: Unsupported or unbounded configuration
+- **WHEN** a profile name, packet count, queue bound, loss, reorder, or bandwidth step falls outside the fixed catalog
+- **THEN** the harness rejects it before allocating or running traffic
+
+### Requirement: Fairness and cap qualification
+The harness SHALL exercise the production fair pacer with eight equal-tier synthetic sessions for the required 60-second virtual interval, retain every per-flow and aggregate observation, report every share error and Jain's fairness index, and fail above 10% share error. It SHALL apply 25% and 50% capacity steps, measure convergence of observed allocation rather than first delivery, fail convergence beyond ten virtual seconds, and fail aggregate egress above 105% of the cap over any rolling five-second window.
+
+#### Scenario: Equal-tier and capacity-step evidence
+- **WHEN** the frozen candidate runs scheduler qualification
+- **THEN** the artifact contains raw per-flow bytes, aggregate-cap series, share errors, Jain's index, measured allocation convergence, and rolling cap observations derived from the production pacer
+
+### Requirement: Honest qualification boundary
+Qualification artifacts SHALL contain no provider endpoint, credential, clipboard text, input payload, secret, or raw media content and SHALL make no claim of live Apollo/macOS/firewall interoperability. The harness SHALL add no codec operation, production dependency, cgo, sidecar, direct provider route, or duplicate processing/impairment simulator. Deterministic smoke evidence SHALL remain distinct from the single normative run on the frozen immutable consumer candidate.
+
+#### Scenario: Deterministic evidence publication
+- **WHEN** qualification completes
+- **THEN** the manifest labels fake-provider, path impairment, and local processing evidence separately and leaves live interoperability deferred-owner-e2e
@@ -0,0 +1,19 @@
+## ADDED Requirements
+
+### Requirement: Provider terminal events end forwarding
+Encrypted provider termination and unexpected provider disconnect SHALL quiesce provider ingestion and queued/new media forwarding before the existing reliable typed terminal event is delivered, close the Verse tunnel within a bounded interval, release the session reservation, and report the appropriate durable provider/session state. A fixed drain delay MUST NOT stand in for reliable control delivery.
+
+#### Scenario: Host termination closes the tunnel
+- **WHEN** the native provider emits an authenticated termination event
+- **THEN** queued and newly injected media cannot cross the Verse transport after observation, and the client tunnel, reservation, and durable lifecycle transition complete
+
+#### Scenario: Unexpected provider disconnect is reconnectable
+- **WHEN** required provider transport disconnects without acknowledged termination
+- **THEN** forwarding stops and the Server receives the existing reconnectable lifecycle state rather than a termination claim
+
+### Requirement: Cleanup failure remains durable
+Gateway cleanup MUST preserve `cleanup_pending` when provider input release, transport cleanup, authorized cancellation, or durable reporting fails.
+
+#### Scenario: Terminal cleanup fails
+- **WHEN** a provider terminal event is handled but required cleanup cannot complete
+- **THEN** the session is not reported released or reusable and durable state remains cleanup pending
@@ -0,0 +1,24 @@
+## 1. Native Policy and Media
+
+- [x] 1.1 Drive supported immutable policy values into encrypted Apollo ANNOUNCE
+- [x] 1.2 Negotiate registered profiles through shared Protocol intersection behavior
+- [x] 1.3 Reject unsupported client or provider/source policy before `/applist`, `/launch`, or readiness
+- [x] 1.4 Evict oldest incomplete audio FEC state and pass sustained-loss relay regression
+
+## 2. Lifecycle and Telemetry
+
+- [x] 2.1 Quiesce queued and new media before terminal event delivery and report durable state
+- [x] 2.2 Preserve cleanup-pending on terminal cleanup failure
+- [x] 2.3 Report measured heartbeat egress and separately sampled queue, processing, and pacing telemetry
+
+## 3. Qualification Path
+
+- [x] 3.1 Add a red-to-green native UDP-to-public-client production-path smoke gate with per-traversal stage evidence
+- [x] 3.2 Remove direct native internals, private QUIC/parser, and arithmetic impairment shortcuts
+- [x] 3.3 Retain raw processing, impairment, fairness, cap, convergence, queue, and resource observations
+- [x] 3.4 Pass short fixed-profile, impairment, fairness, race, parser fuzz, and resource smoke checks
+
+## 4. Immutable Freeze
+
+- [x] 4.1 Pin and verify a separately published never-reused Protocol version from an empty cache
+- [x] 4.2 Freeze all normative inputs and run the corrected Section 7 qualification exactly once