Phase 7 — hardening, load test & launch (v0.1.0)

[MANVENDRA-github]

Phase 7 — Hardening, Load Test & Launch

The final phase: turn a feature-complete gateway into a launch-ready, v0.1.0 artifact. Three logical commits — security hardening, a measured load harness, and launch hygiene.

Part B — Security hardening (cfa9477)

Closes the launch-blocking gaps in SECURITY_REVIEW_LOG.md:

• Per-API-key inbound rate-limiting (CLIENT_RPM) via a per-client token bucket — one runaway key gets 429s without affecting other clients (box 7.1, new code + tests).
• SSRF hardening: upstream fetch uses redirect: 'error' so a malicious provider can't 3xx-redirect the prompt to another host (box 4.2).
• Log redaction: exported logRedaction + a test proving the authorization header logs as [redacted] and the raw key never appears (box 1.2).
• CI dependency audit: pnpm audit --prod (prod deps clean; dev-only advisories triaged, box 8.1).
• Ticked boxes 3.2 / 6.2 / 8.2 from existing coverage; added SR-006 and completed the pre-launch gate.

Part A — Load harness + measured numbers (8cd3c1a)

pnpm load — an in-process harness (two mock upstreams + the gateway via Fastify inject) that exercises the cache, fallback, and guardrail paths and writes load/RESULTS.md. It's excluded from lint/typecheck and run manually, not in CI. The mock serves a deterministic /embeddings endpoint so exact-repeat prompts truly hit the semantic cache. This fills the long-standing X/Y placeholders in PRP_SPEC.md §5 with real measured numbers:

Metric	Result
Cache cost-reduction (50%-repeat workload)	50% — 100/200 served from cache, zero upstream calls
Unhandled 429s to the client	0 — all 50 always-429 requests retried + failed over
Guardrail catch-rate (injected PII)	100% — 50/50 blocked in-path
Added overhead (Sentinel's own time)	~14 ms p99 (p50 ~7.5 ms) vs a near-instant mock

Framing is stated honestly in load/RESULTS.md and the README: overhead is Sentinel's own per-request cost (not a model's latency); cost-reduction tracks your traffic's repeat rate; the catch-rate is the deterministic guardrail rate (the async LLM judge needs a real model and is covered by unit tests).

Part C — Launch hygiene (37d23fb)

• LICENSE (MIT) and CHANGELOG.md (v0.1.0, grouped by capability area).
• docs/dashboard.png — a Playwright-generated screenshot on seeded data (the screenshot spec reuses the E2E route-stub pattern and doubles as a render smoke test).
• README polish: CI + license + node badges, a current-state v0.1.0 banner replacing the stale "Phase 1" header, a Benchmarks section, the embedded dashboard screenshot, and License + Contributing sections.

Verification

• pnpm verify green — typecheck + lint + 203 tests at 98.6% statements / 92.6% branches (above the 90% gate). load/** is lint/typecheck-excluded; the new screenshot spec lints clean.
• pnpm test:e2e green — both dashboard specs pass; docs/dashboard.png regenerates.
• pnpm load runs clean and prints the metrics table above.

Follow-up (post-merge)

• Tag v0.1.0 + gh release create from the CHANGELOG notes.

Closes out the ROADMAP. After this, Sentinel is feature-complete, hardened, benchmarked, and documented.