perf(server): reduce middleware pipeline overhead - #1950

✅ No new issues found.

I reviewed the whole diff and traced the pipeline logic line-by-line. The change is a clean, well-scoped refactor with no functional regressions.

Reviewed changes

• Per-procedure caching: Two WeakMaps keyed on the procedure — getProcedureCallArtifacts caches the error-constructor map (which allocates a Proxy) and reconcileError; getExecutionPlan precomputes the per-level schema slice boundaries, the stacked-object snapshot, and validation flags once instead of per request.
• Untraced fast path: getOpenTelemetryConfig()?.tracer is read once per call; when no tracer is configured the middleware/validators/handler run directly, skipping the runWithSpan wrapper. Span names and attribute side effects are unchanged when traced.
• Conditional context copy: hasEnumerableProperties decides whether to spread-merge or reuse the same context object when a middleware adds no data; symbol keys are handled. Syntactic equivalence to the old always-copy path holds for the merged content.
• Micro-opts: primitives skip the instanceof ORPCError branch, non-lazy procedures skip the unlazy resolved-promise, and next context extraction reads rest[0]?.context directly.
• Tests & benches: new tests meaningfully exercise the empty-{}/next(undefined)/symbol-key context cases, the context: undefined result, primitive outputs, and the exact traced span-name sequence. Benches extend existing files. Full test file passes (44/44).

ℹ️ Middleware context now aliases across pass-through frames

This is a documented, deliberate tradeoff, not a bug — but it's slightly subtler than "the next level gets the same object". Previously every next() performed { ...context, ...nextContext }, which gave each middleware frame an isolated copy. On the new path, when a middleware passes no context data the identical object reference flows down, so an in-place mutation of options.context inside a middleware is now visible to ancestor frames that previously received their own copy. The content-equality contract is intact; only in-place mutation aliasing changes. Given use() snapshots schemas into immutable OrderedMiddleware entries and contexts are conventionally read-only, this looks safe — just flagging it as the one caveat reviewers should keep in mind rather than a change to request.

# Context object aliasing across pass-through middleware frames

## Affected sites
- packages/server/src/procedure-client.ts:379-389 (the middleware `next` callback: `hasEnumerableProperties` decides merge-copy vs. reuse)
- packages/server/src/procedure-client.ts:406-409 (result context reuse)

## Required outcome
- No change required. Content-merge semantics are preserved; only object identity differs from the previous always-copy behavior.
- Document, if not already, that middlewares must not mutate `options.context` in place and rely on isolation, since passthrough frames now share the same object.

  ^{｜ View workflow run ｜ Using DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏}

✅ No new issues found.

Reviewed changes

Since the prior pullfrog review (466cff4), this run reviewed commit 5eead6e:

• Added a test covering the untraced input/output validation fast path — a transformed .input(...)/.output(...) pipeline is exercised with the no-op tracer disabled, so the test would fail if the fast path skipped transform application.
• Simplified hasEnumerableProperties by dropping the now-redundant undefined early-return and narrowing its parameter to Context; both call sites (nextContext !== undefined && … and resultContext !== undefined && …) already guard against undefined before invoking.

The narrowed signature is type-safe and behavior-neutral, and the new test meaningfully exercises the branch it claims to cover. The full test file passes (45/45).

  ^{｜ View workflow run ｜ Using DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏}