[LarsArtmann]

What

Evaluations now report step-level progress instead of an opaque spinner:

• judge.Options.OnProgress receives milestone events (start, rubric skip/reuse/generate/complete/fail, scoring, retry, done, error); nil keeps the CLI behaviour unchanged
• schema/progress.schema.json documents the event contract, with a schema-parity test
• the server keeps a per-job append-only progress log and exposes GET /api/sessions/{id}/analyze/stream: SSE progress events with id: lines, then a terminal status event. Last-Event-ID reconnects resume where they left off
• Start now uses http.Server with a ReadHeaderTimeout (Slowloris cap, gosec G112) and gains Shutdown; serve/open/map handle SIGINT/SIGTERM with a 5-second drain so in-flight SSE handlers finish instead of dying mid-stream

Why

A judge run takes one to two minutes with zero feedback; the SSE contract makes the pipeline phase observable by any client (the browser UI consumes it in my fork; the API is useful on its own and the schema pins it).

Series

Part 8/9; stacks on #31 (and #25–#30 before it). "Files changed" shows the cumulative stack — this PR's own delta is just the eighth commit, feat(judge+server): stream judge progress over SSE with graceful shutdown.

Validation

• go test -count=1 ./... and go test -race ./internal/server/ green at the tip
• new tests: SSE framing + resume offset, progress schema parity, port-0 binding, Shutdown-before-Start no-op