Measure scaling and bound every host

01 / CompilerRead scaling and size as separate budgets

Version 0.4 has no default component or connection count. Note that this is a compiler capability, not permission for a public endpoint to accept arbitrary work — the two are easy to confuse and expensive to confuse. Whenever the source is not ours, set components, connections, sourceCharacters, wireCrossings, and svgOutputBytes through the per-compilation limits option.

A wall-clock deadline is still required on top of that, because every one of those fields bounds a size and none of them bounds time. A small document with pathological geometry can be expensive to route, and no count of output bytes will catch it.

port:PA "A" at (70, 100) #cyan
resistor:RP "R" at (330, 100) #amber [orientation=down]
component:SHLD "shield" at (330, 250) #slate [width=150 height=70]
port:PB "B" at (700, 100) #emerald

PA.out -> RP.in #cyan [ortho]
RP.out -> PB.in #emerald [ortho]

The release gate measures two different client costs. A tree-shaken host that imports only compileSchematic must remain below 35 KiB gzip; the complete public entry, which includes exports a given host may never call, must remain below 37 KiB gzip. Both are budgets, not claims that TypeScript declarations, the npm tarball, installed files, or a website route have the same size. Run bun run size on the exact release commit instead of repeating a registry screenshot as a bundle measurement.

The compiler budget rose from 34 KiB in 0.7, and it is worth being explicit about why rather than quietly restating a number. Bundle nudging cannot tree-shake out of the compile path — the router reaches it, so every document that routes carries it — and it is what takes a reversal bus from twelve wires to thirty-two. The cost is recorded against the capability that bought it, in the same terms 0.5 used when relative placement moved the budget from 32. Four separate attempts to buy the bytes back are recorded in the changelog: packed encodings, glyph tables, literal hoisting and family partitioning each made the gzipped bundle larger or returned far less than projected, because gzip has already done that job better than a hand-packed encoding can. On this package, optimising for minified size is optimising against the metric that ships.

Scaling is gated independently. The 0.4 release measured 64,000 components with 32,000 connections at roughly one second and a flat cost near 16 µs per component from 8,000 upward; 11,200 components with 800 obstacle-dodging orthogonal traces measured about 130 ms.

0.7 moved the benchmark medians rather than the slope. A 512-component figure compiles in 2.69 ms and dense 16×16 orthogonal routing in 4.18 ms, both from a run where the family predicates stopped answering by linear scan — four of them ran KINDS.includes(kind), and the false answer is the expensive one, so every electrical component walked all forty-seven UML keywords before being told no. A twelve-wire contended bus compiles in 25.2 ms, from a spatial-bucket query that no longer allocates an empty array on a miss in the innermost of three nested loops.

Those are warm medians from one Apple Silicon/Node configuration, and the durable claim is the guarded slope rather than the absolute millisecond value. Run bun run benchmark on our own hardware before quoting any of them.

02 / HostIsolate deadlines and cap amplification

The official compile endpoint applies limits stricter than the library defaults, rejects oversized request bodies before JSON allocation, rate-limits by client, and runs a compilation outside the request thread. A deadline can terminate that worker; Promise.race around synchronous compilation cannot. Concurrency and queue depth are finite, and a full queue returns a retryable capacity error instead of multiplying memory by the rate-limit burst.

clock:CQ "CLK" at (90, 120) #purple
flipflop:FQ "FF" at (340, 120) #cyan [type=d]
port:QO "Q" at (620, 120) #emerald

CQ.out -> FQ.clock #purple [digital ortho]
FQ.out1 -> QO.in #emerald [digital line]

Successful and line-numbered failed compilations share a SHA-256-keyed LRU capped at 64 entries and 16 MiB. The cache key includes mode, bounds, title, and source; hits refresh recency. Registry data uses one immutable stale-while-revalidate snapshot and one in-flight refresh promise. Documentation and simulation caches are keyed by finite registries rather than user input.

The browser IDE compiles locally in a lazily started native worker containing @schemd/core and uses the endpoint only as a fallback. That removes a network round trip and server work from normal typing, and the worker boundary lets a deadline terminate a pathological local compile, without putting the compiler on documentation, catalogue, or simulation critical paths. The website build budget verifies both facts from Vite’s manifest.

Performance work is not finished when one benchmark turns green. Re-run bun run benchmark on the target hardware, read p95 and p99 under concurrent load rather than a warm median, and measure the serialized response — not only parser time — before raising any public-host budget.