eval: add explicit design then implementation workflow

[iantocristian]

Summary

• adds an eval-only --workflow design-implement mode to the PR replay runner
• runs a read-only structured design phase, verifies it produced no patch, then injects the captured design verbatim into a fresh implementation call
• records prompts, responses, event streams, token/command usage, duration, and jscout requests separately per phase and in aggregate
• joins phase-specific telemetry correctly and refuses report-level pooling across single-phase and two-phase runs
• records G15/G16 as parked product work; PR G15: add design-before-edit task memory #45 remains open with the blocked label

This adds no jscout command, MCP tool, semantic artifact, skill guidance, or runtime product behavior.

Failure behavior

A timed-out, nonzero, invalid, or source-mutating design phase does not unlock implementation and is not graded as an implementation attempt.

Validation

• npm test (57 tests)
• node --check scripts/eval-run-replay.mjs
• git diff --check

[iantocristian]

Replay report added in commit 0a6588e: https://github.com/iantocristian/jscout/blob/codex/eval-two-phase/eval/results/next-root-layout-param-types-2026-08-18.md

The valid matrix completed 12 arms: 8/12 passed. The host-sleep Terra timeout is preserved but excluded and replaced by an unchanged retry that passed. Main findings: incremental bounded jscout usage worked; exact-identifier ranking and semantic-memory selection did not; memory-profile 4/4 is not attributable to retrieved artifacts; design-then-implement cost more and passed 3/6 versus 5/6 single-phase; dev self-verification still exposes EMFILE and sandboxed-install EPERM.

[iantocristian]

Follow-up pushed in d6c3014.

Correction: unconditional Playwright sidecar startup was introduced by 847148b and is already on main; it was not introduced by PR #46.

This PR now makes that existing behavior capability-aware without changing launcher contracts:

• browser_server defaults to auto, starts only when the prepared workspace resolves Playwright;
• required fails closed if Playwright is missing or launch fails; disabled skips explicitly;
• per-arm browser-server.json and response rows record the decision;
• Next replay task sets use required and WATCHPACK_POLLING=250;
• the agent test command includes NEXT_SKIP_ISOLATE=1 to prevent nested installs in the sandbox.

Parallel runners need no invocation changes: ports are OS-assigned and registries/workspaces remain per arm/process. Existing output roots/trial IDs still need to be unique, as before.

Local verification: npm test — 62 passed, 1 platform skip, 0 failed.

[iantocristian]

Plan/eval follow-up pushed in 6d70ee7.

• G15 remains parked after the root-layout two-phase arms preserved wrong contracts.
• G16 is now explicitly independent of G15 and limited to G14 attached-memory false negatives; the root-layout run does not trigger it because the corpus lacked a useful causal artifact.
• G17 is the immediate exact-identifier dominance milestone.
• G18 covers scope-stratified/targeted scouting and support-aware compact semantic-memory retrieval without larger budgets.
• Both committed Next task sets now inject NEXT_SKIP_ISOLATE=1 alongside WATCHPACK_POLLING=250.

Validated task JSON, diff whitespace, and replay/report tests: 15 passed.