Whole-repo hardening round: streaming, content policy, realtime billing, modernization

Fixes from a full-source review:
- streaming responses outlived the per-account timeout: reqwest's request
  timeout is a total deadline, so any generation longer than timeout_seconds
  (default 60s) was cut mid-stream and billed as aborted; streams now get the
  bound on response headers and then per gap between chunks
- REST responses never masked secrets (detect_secrets applied inbound and on
  the realtime relay only), and a secrets-only tenant streamed raw deltas;
  outbound masking and the stream-buffering boundary now share one predicate
- the blocklist scanned base64 image parts, randomly blocking multimodal
  requests; the scan now visits the same text fields DLP does
- a second realtime response.create during an active turn desynced the
  admitted-turn queue for the rest of the session (misattributed billing plus
  a stuck reservation); the bridge holds at most one admitted turn
- content-blocked requests on the responses/embeddings/images/audio surfaces
  answered 500 instead of a graceful 400
- vendor error bodies without an error envelope (Bedrock, DashScope native)
  parsed as empty successes on 4xx/5xx statuses
- claude engine joined system texts without a separator
- MemoryStore's ledger prune wedged behind an out-of-order young row
- config generation hashed with DefaultHasher (not stable across rustc), so a
  mixed-build fleet would miss the shared cache; sha256 now, builder pinned
- Redis QPS used ceil() where the in-process limiter rounds; negative
  quotas/qps/qpm are rejected at config load; the in-process failure streak
  decays after an hour idle like the Redis backend's TTL

Modernization and hygiene: rust-version 1.88 (let-chains were already in use),
resolver 3, unmaintained serde_yaml replaced by serde-saphyr, yanked spin
bumped, account slots shared via Arc instead of cloned per request, the
one-caller response-transform DSL replaced by a direct conversion, SSE decoder
and buffered drains bounded against unbounded peers, OpenAI/Anthropic usage
now carry cache/reasoning detail on the wire, docs re-synced to the code, and
regression tests pin every fix above plus an e2e erasure round trip.

Verified: 329 tests green on macOS and on Linux (rust:1.94, live PG/Redis);
bench A/B on bare metal: small-chat p50 and request-clone unchanged,
concurrent throughput up ~3%, big-payload p50 +14% traced to x86 codegen
drift (bisect halves it with no per-byte change present; arm64 shows parity).

Author: CMGS

Cargo.lock

@@ -1,6 +1,6 @@
 # Gateway — LLM access point. Workspace manifest.
 [workspace]
-resolver = "2"
+resolver = "3"
 members = [
     "crates/consts",
     "crates/models",
@@ -20,7 +20,7 @@ version = "0.1.0"
 edition = "2024"
 license = "AGPL-3.0-only"
 repository = "https://github.com/cocoonstack/gateway"
-rust-version = "1.85"
+rust-version = "1.88"
 publish = false
 
 [workspace.lints.clippy]
@@ -43,7 +43,7 @@ reqwest = { version = "0.12", default-features = false, features = ["rustls-tls"
 # --- serde ---
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
-serde_yaml = "0.9"
+serde-saphyr = "0.0.29" # serde_yaml is unmaintained (archived 2024)
 # --- errors ---
 thiserror = "2"
 anyhow = "1"

Cargo.toml

@@ -1,4 +1,6 @@
-FROM rust:1 AS builder
+# Pinned: successive image builds must embed the same toolchain, not whatever
+# `rust:1` floats to on the day of the build.
+FROM rust:1.94 AS builder
 WORKDIR /app
 COPY . .
 RUN cargo build --release -p gw-server --locked

Dockerfile

@@ -54,7 +54,7 @@ docker run -p 8080:8080 -v $PWD/conf/gateway.yaml:/etc/gateway.yaml \
 ```
 
 The image binds `0.0.0.0` (`GW_HOST`) and ships a `/health` HEALTHCHECK.
-Published multi-arch to `ghcr.io/cocoonstack/gateway` on push.
+Published multi-arch to `ghcr.io/cocoonstack/gateway` on `v*` tags.
 
 ## Development
 
@@ -64,12 +64,13 @@ make test     # cargo test --workspace
 make lint     # clippy -D warnings
 make fmt      # cargo fmt --all
 make deny     # cargo deny check (advisories + licenses)
-make release  # optimized gw-server binary (--locked)
+make release  # optimized `gw` binary (--locked)
 make docker   # build the container image
 ```
 
-CI runs fmt/clippy/test + `cargo deny` on every push; tagged `v*` pushes
-build multi-arch binaries (release) and a multi-arch image (docker).
+CI runs fmt/clippy/test + `cargo deny` on every push to `main` and every PR;
+tagged `v*` pushes build multi-arch binaries (release) and a multi-arch image
+(docker).
 
 ## License
 

README.md

@@ -9,7 +9,8 @@ description = "Local file configuration — no config center."
 [dependencies]
 gw-consts = { workspace = true }
 serde = { workspace = true }
-serde_yaml = { workspace = true }
+serde-saphyr = { workspace = true }
+sha2 = { workspace = true }
 thiserror = { workspace = true }
 regex = { workspace = true }
 tracing = { workspace = true }