fix: harden the capture pipeline for production use

[RubenGlez]

Fixes 21 correctness, performance and reliability issues found while auditing the library for real-world usage. Each fix ships with a regression test.

Silent data loss and hangs

Hono + Next.js App Router lost every request body	Both read the body after the route handler had consumed it. Cloning a consumed Request throws, the catch swallowed it, and requestBody was recorded as null on every POST/PUT/PATCH. The existing Hono test passed only because its handler ignored the body.
Streaming responses hung the client	Hono, Next.js and Elysia awaited .json() on a clone of the response before returning it. On an open SSE stream that never resolves, so the client received nothing.
Offline mode never worked	The Ollama provider used the AI SDK's default OpenAI model, which targets /v1/responses. Ollama (and most OpenAI-compatible gateways) only implement /v1/chat/completions, so every generation failed — breaking both privacy.offline and the no-API-key fallback.
The circuit breaker never reopened	Five consecutive failures disabled capture for the process lifetime, so a transient provider outage silently stopped all documentation until the next deploy. Now retries after a 60s cooldown.
NestJS never documented errors	The interceptor used tap(), which only fires on success — every response from an exception filter went unrecorded.
Next.js Pages Router exploded endpoints	Route params weren't separated from the query string, so /api/users/1 and /api/users/2 became separate rows, each paying for its own AI call.

Robustness

• capture() can no longer throw into the host app's response path, and self-referential bodies terminate instead of overflowing the stack.
• Concurrent workers (cluster mode, multiple replicas) no longer collide on the unique endpoint index after the AI call has been paid for — project and endpoint writes are atomic upserts.
• Capture is bounded by default: 256 KB bodies, non-JSON responses skipped, 50 shapes per endpoint, 50 spec versions per endpoint. Previously maxBodySize had no default, so a stalled provider could let the queue retain unbounded payloads in the host app's heap.
• Capturers expose flush(), drained on shutdown where the framework has a hook (Fastify onClose, Elysia onStop, NestJS onApplicationShutdown).
• Default models updated to current, non-retired IDs; a rejected model now produces an explicit "pin ai.model" error rather than an opaque 404.

Performance

• Repeated payload shapes are dropped synchronously, before the queue, so steady-state traffic neither retains bodies nor queues behind an in-flight generation.
• The queue processes endpoints in parallel (still one shape at a time per endpoint), instead of one capture at a time globally.
• The size cap stops as soon as the limit is exceeded rather than serializing the whole payload (~2ms → microseconds on a 1 MB body). Privacy rules are compiled once per config instead of per request.
• The dashboard reuses one database handle; three API routes were opening a new client and re-running the schema DDL on every request.
• Request bodies are trimmed like responses before going to the model, so a bulk payload is no longer billed in full.

Breaking

The Postgres helpers (createPgDB, pgGetAll, …) moved from the package root to @easydocs/core/storage/postgres, and postgres is now an optional peer dependency loaded on demand — the root re-export pulled the driver into every SQLite install. Postgres users must npm i postgres. Configuring storage.type: 'postgres' is unchanged.

Verification

• build, lint, typecheck green.
• 254 tests pass, up from 202. @easydocs/nextjs had no test setup at all and now has 7.
• End-to-end run of real Express and Hono apps against a stub AI provider, asserting: request bodies reach the model, dynamic ids collapse to one endpoint, the SSE route returns in 0ms and creates no endpoint, repeated shapes cost exactly one generation, and no raw secret is stored.

Not addressed

pnpm audit --prod reports 2 pre-existing high CVEs (js-yaml, next > postcss > nanoid), both in apps/dashboard dependencies and untouched by this branch. They will block pnpm release until bumped.

https://claude.ai/code/session_01GrfumUQ4vFUAgsFYwzDyK8