Priority exploration and a trace timeline

[dhruvl]

Two additions on one substrate: schedules can now be explored by priority with a stated per-run guarantee, and any run's trace can be drawn as a timeline you can open in a browser.

• The trace grew the substrate both need, in one disclosed format change: every scheduler event now names the host whose code it runs (empty for the simulation's own steps), and the network records a deliver event paired to each send by uid. Recorded trace hashes therefore differ from 0.1.0 for every workload — the changelog says so plainly. TraceEvent is now a named tuple; recording is slightly cheaper than before.
• Scheduling policies now see who is ready, not just how many: the policy seam receives a lazy sequence of (owner, label) views, where owners are per-loop task creation indices. The default seeded policy and scripted replay consume only the count, and their schedules are byte-identical to before — verified across seeds against the previous build.
• policy="pct" runs Burckhardt et al.'s Probabilistic Concurrency Testing (ASPLOS 2010): distinct random priorities per chain of work, the highest-priority ready entry runs, and depth-1 change points demote the leader. The docs state the honest trade: on a shallow planted race uniform random finds the bug ~46x sooner; what PCT buys is a stated per-run floor for depth-d bugs, which uniform draws cannot promise. The horizon calibrates from seed 0's step count; PCT-found failures replay by seed and by recorded choice list.
• --simloop-timeline renders each failing seed's trace to a self-contained HTML page: one lane per host plus one for the simulation, send-to-deliver arrows (duplicates, held-then-released re-sends and transmit-time holds all drawn honestly), crash and restart marks distinct from packet fates, hover details, zoom and pan, and a truncation banner past five thousand events. No external requests of any kind.
• Benchmarks re-measured on this tree and republished: the single-process chaos campaign now explores ~52 seeds/second, down from ~55 in 0.1.0 — the per-delivery trace events are a feature with a cost, and the numbers say what they measured.

Suite green (455 tests + slow markers); strict mypy clean; no new runtime dependencies.