feat(server): harden hostile input and overload behavior (#15)

docs/security/abuse-protection.md

@@ -0,0 +1,98 @@
+# Hostile-input and overload protection
+
+Tracking: #15
+
+Rendezvous treats every public HTTP request and UDP datagram as hostile. The
+server applies bounded fixed-window request budgets and concurrency ceilings in
+two stages so malformed input is discarded before expensive work while valid
+traffic is also isolated by its authenticated scope.
+
+## Enforcement order
+
+1. Kestrel and the HTTP abuse middleware cap request bodies at 16 KiB. A known
+   oversized body receives a typed `413` response before endpoint dispatch.
+2. Every HTTP request consumes global, source-prefix, and operation budgets and
+   acquires the corresponding concurrency leases. IPv4 sources share a `/24`
+   budget and IPv6 sources share a `/56` budget; raw addresses are not retained.
+   Non-lease operations also consume a smaller optional-work budget, leaving a
+   configured global and source-prefix reserve for renew, update, and delete
+   operations during shedding.
+   Health probes use their own source-prefix budget so public API overload cannot
+   make a healthy instance fail its orchestrator probes, while health traffic is
+   still bounded.
+3. Once an endpoint has safely derived identities, it also acquires applicable
+   tenant, principal or capability, and listing/attempt budgets. Secret
+   capabilities are represented only by bounded SHA-256 fingerprints.
+4. Every UDP envelope consumes global, source-prefix, and wire-operation
+   budgets before decoding. A structurally and cryptographically valid request
+   then consumes capability, role, and mediation-handle budgets before state
+   mutation or introduction.
+5. HTTP overload returns the stable `RateLimited` error, status `429`, and a
+   bounded `Retry-After` value in both the header and response contract. UDP
+   overload and every invalid UDP input are silently dropped.
+
+The same HTTP identity budget is computed whether or not a listing or attempt
+exists. Rejection therefore does not disclose resource existence. Publisher
+authentication also completes before any tenant/resource operation, while the
+pre-authentication source budget prevents invalid credentials from bypassing
+load shedding.
+
+## Bounded state and recovery
+
+`Rendezvous:AbuseProtection:MaxTrackedKeys` is a hard combined ceiling for rate
+and active-concurrency keys. General HTTP and UDP traffic cannot consume the
+configured `CriticalTrackedKeyReserve`; lease operations and health probes may
+use that reserve but never exceed the hard ceiling. A request that would exceed
+its applicable ceiling fails closed without adding state. Fixed-window rate keys
+are cleared at the next window boundary; concurrency keys are removed as their
+request leases finish. HTTP and UDP trackers have separate locks and cardinality
+partitions, so a UDP flood cannot block HTTP admission on a shared lock or
+consume HTTP key capacity. This gives
+deterministic burst recovery and prevents an attacker from growing a permanent
+high-cardinality address, credential, or resource table.
+
+The complete default profile is checked into
+`src/FinalFactory.Rendezvous.Server/appsettings.json`. Operators may lower or
+tune limits for a measured deployment profile, but must preserve all dimensions
+and leave the tracker ceiling above the maximum simultaneous key set. A rolling
+deployment should use the same profile on every instance. These per-process
+limits are a final service boundary; an edge proxy may add stricter distributed
+limits but is not a substitute for them.
+
+When an HTTP reverse proxy is used, every immediate proxy address must be
+allowlisted in `Rendezvous:AbuseProtection:TrustedProxyAddresses` (or indexed
+environment variables such as
+`Rendezvous__AbuseProtection__TrustedProxyAddresses__0`). Only one forwarded
+hop is accepted. With an empty allowlist, forwarded headers are ignored and the
+direct TCP peer is the source. Never add a broad network range or accept
+untrusted `X-Forwarded-For` input: that would let a caller choose its own rate
+partition.
+
+## Reflection, disclosure, and logging rules
+
+- UDP sends nothing for malformed, oversized, unauthenticated, stale,
+  replayed, wrong-role, or rate-limited input.
+- Introductions are emitted only after both role-scoped capabilities bind to
+  their observed gameplay-socket sources. HTTP never supplies a public
+  introduction target.
+- Private candidates must be same-family private unicast addresses and are used
+  only for peers observed behind the same public address.
+- Abuse keys, exceptions, and responses never include bearer credentials,
+  capabilities, tickets, raw endpoints, metadata values, or hostile markup.
+- Endpoint and capability values are not used as metric labels or log fields.
+
+## Verification
+
+The deterministic test corpora use the recorded seeds `0x152026`, `0x154A50`,
+and `0x1557A7E`. They exercise 10,000 arbitrary UDP envelopes through the
+production decoder, 5,000 arbitrary HTTP/credential parser inputs, and 1,000
+mutated state transitions, including the oversized and configured-capacity
+boundaries.
+Focused tests cover IPv4 and IPv6 prefix
+partitioning, tenant/principal/resource concurrency, tracker exhaustion,
+window recovery, wire-operation isolation, a steady-state allocation ceiling,
+typed `429`/`413` responses, secret fingerprint redaction, and silent
+authenticated UDP shedding. The existing state, contract, HTTP, client,
+and mediator suites continue to cover cross-tenant access, replay, role swaps,
+credential rotation, bounded metadata, endpoint validation, and one-shot
+amplification behavior.

docs/security/control-matrix.md

@@ -11,7 +11,7 @@ backlog where the control is implemented and verified.
 | Per-game credentials and signing keys | Provisioned principals and versioned keys are scoped to game/environment; secrets come from a provider and never a public binary. (#5) | Cross-tenant authorization tests, rotation/overlap/revocation tests, and secret scans. |
 | Short-lived, single-purpose tokens resistant to replay | Issuer fixes audience, tenant, attempt, role, issued/expiry times, nonce, and key ID; store atomically consumes nonce/ticket. (#4, #6, #10) | Golden vectors; expired, future, mutated, wrong-role, wrong-tenant, and concurrent replay tests. |
 | Strict payload, metadata, and token size limits | ADR 0003 ceilings are checked before allocation/deserialization and again at domain construction. (#4, #15) | Boundary/property tests, malformed corpus, and allocation-aware fuzzing. |
-| Registration, query, and introduction rate limits | Layered per-address, principal, tenant, and global token buckets with bounded queues and stable retry guidance. (#15) | Limit partition/isolation tests and overload/soak profiles. |
+| Registration, query, and introduction rate limits | Layered fixed-window budgets and concurrency leases cover global, operation, IPv4 `/24` or IPv6 `/56`, tenant, principal/capability, and listing/attempt dimensions with a bounded key table and stable retry guidance. (#15) | Deterministic partition, concurrency, tracker-exhaustion, recovery, typed-overload, and silent-UDP-shedding tests. |
 | Lease expiry removes abandoned servers | Visibility and join eligibility atomically require a fresh lease and fresh authenticated presence. (#6, #7) | Fake-clock expiry, renew/expire race, restart, and stale-host join tests. |
 | Validate game, environment, room, and protocol boundaries | Every identifier is a validated type; store keys and authorization decisions include server-derived tenant scope; protocol is exact-match in v1. (#4-#10) | Contract, tenant-isolation, incompatible-version, and confused-deputy tests. |
 | Structured audit events without secrets or reusable credentials | Allowlisted audit schema excludes metadata values, raw endpoints, tokens, and key material; event volume is bounded. (#16) | Captured-log/audit assertions and credential canary scans. |