Design: upstream jco node/browser driver layer (execution topologies survey)

[lann]

Deferred half of the consumer-glue consolidation (see the bin/helpers issue and umbrella on #14). Each consumer built a jco driver over @polymorph/component-test-js/harness, and the topologies genuinely differ:

• webcrypto (conformance/driver-ct/jco/runner.mjs, worker-node.mjs, run-browser.mjs, worker-browser.mjs): worker-thread / Web-Worker pools striping ~19k cases, fresh instance per case (JSPI attempts cannot be cancelled), content-stamp + flock transpile guard, Playwright page harness with heartbeat stall watchdog, findChrome() ladder.
• websocket (run-node.mjs, run-browser.mjs, harness.mjs): sequential (loopback-I/O-bound corpus; sidesteps Chromium per-endpoint handshake serialization), suite-owned echod spawn, env-carried config.
• webrtc (run.mjs, run-browser.mjs, harness.mjs): drivers are children of a Rust orchestrator — JSONL on stdout, role/run-id env contract, --select prefix, same-origin reverse proxy for the browser mailbox.

What is provably common: transpiled-module loading, import binding, envelope emission, per-case timeout handling, the browser page-driver skeleton (static server + import map + Chromium + watchdog). What varies: concurrency/instance policy, SUT host modules and their config, child-process vs in-process contract.

Decision deferred until (a) the shared-parts issue lands and shrinks the residue, and (b) a fourth consumer arrives (tls's jco leg, blocked on lann/jco#34 + lann/jco#35) to test any proposed seam against. Generalizing the driver from n=3 with live divergence risks an abstraction the next consumer immediately breaks.

[lann]

The survey's precondition is met: the fourth consumer exists (polymorph-tls's jco-node leg, tls#33), and it re-derived the same sequential driver skeleton the other three carry — the evidence the deferral asked for. The node half lands via the PR above (runSuiteJsonl + ./node-runner); consumers adopt on their next pin bump. The browser half (page driver: static server, import map, worker/page pool, stall watchdog, findChrome ladder) remains this issue's open scope, now with a named first consumer waiting: polymorph-tls#32 — seven loopback cases, no SUT host, no env; the simplest possible seam test, same forcing-function pattern that validated #58.

[lann]

Browser half landed (#71, keyed-by-target fix in #72) and its first consumer is merged: polymorph-tls#35 adds jco-browser/jco-browser-delegated rows on the upstream page driver — the consumer frame came out at ~90 lines (core-URL enumeration + target config + results writing) plus a ~20-line imports module, with the CI leg green on the runner image's Chrome. Remaining before this issue closes: migrate the three existing copies (webcrypto's scripts/browser-page-driver.mjs + worker pool, websocket's and webrtc's run-browser.mjs) onto the upstream driver at their next pin bump.

[lann]

Closing: both halves are upstream and adopted by every consumer.

Node half (#70): runSuiteJsonl + ./node-runner — adopted by tls (#34), websocket (#38), webcrypto (#349, pool keeps composing runCases), webrtc (#145, selection as a tests-interface wrapper).

Browser half (#71, keying corrections #72/#73, launchArgs #74): ./browser-driver (page harness, routes hook, self-mount + import map, Chrome ladder) + the in-page page-runner/browser-worker pair — first consumer polymorph-tls#35 (new jco-browser targets, the seam test), then websocket#39 (env-configured SUT through the generic worker), webcrypto#350 (worker pool, per-entry report keys, custom context; scripts/browser-page-driver.mjs deleted at origin), webrtc#146 (shell-only: RTCPeerConnection is Window-only, so worker-hosted SUTs are a real boundary — the in-page run stays consumer-side by necessity, documented there).

The survey's outcome, versus the original topology table: what proved genuinely common was the loop, the shell, and the shard worker; what proved genuinely per-consumer was SUT/env wiring, concurrency policy, child-process contracts (#62), and — discovered last — Window-only SUT APIs. Every consumer now carries only its frame.

Residue tracked elsewhere: the two-party orchestration contract is #62 (iroh-gated); webrtc's stale-results-on-child-failure recipe hazard surfaced during this adoption and is polymorph-components/polymorph-webrtc-datachannels#147.