feat: run results report CPU time and per-call bridge records from the runtime

[schplitt]

What

Every RunResult (completed / failed / aborted) now reports what the run actually did, measured inside the Rust runtime:

result.durationMs   // wall-clock time of the run (unchanged, now documented as such)
result.cpuTimeMs    // NEW — active V8 execution time, bridge waits excluded (µs resolution)
result.bridgeCalls  // NEW — one record per bridge call attempt, in attempt order:
// [{ name: 'fetch', startMs: 0.4, durationMs: 2.3, argBytes: 180, responseBytes: 41208, ok: true, blocked: false }, …]

• Names are resolved: plain globals as-is, shim stubs under their public name, host-module import leaves as <specifier>.<path> (e.g. tools:search.query).
• Blocked attempts are visible: calls rejected by maxBridgeCalls/maxBridgeCallBytes/function-argument checks get a record with blocked: true — they never reached the host, but the report shows exactly which attempt crossed the limit.
• No payloads, ever — names, timing, and sizes only, so the report is cheap enough to stay always-on.
• One clock: record offsets/durations share the run's clock with durationMs, so entries line up on a timeline; durationMs − cpuTimeMs ≈ bridge waits + scheduling.

Where the data comes from

Recording lives in the Rust runtime (bridge callback records the attempt; the poll loop settles it when the response frame returns). The TS client only decodes the Result frame and maps wire-level stub names to public names — the client measures nothing. cpuTimeMs reads the existing CPU budget through a new µs-resolution accessor (the integer-ms one only fed the 10ms guard and would report 0 for sub-ms runs).

Wire: result frames gain f64 cpuTimeMs + List<BridgeCallRecord> (~40 bytes/record); docs/protocol.md §5.6 documents the layout.

Known gap

Aborted runs report cpuTimeMs: 0 and bridgeCalls: [] — the abort tears the connection down before the runtime can send its result frame. Tracked in #36 (graceful terminate), which matters most for durable-isolates suspensions.

Testing

• Rust (207): records on success incl. metadata/order, blocked record at the limit, CPU ≤ wall, wire encoder layout.
• TS (453): wire decode incl. name resolution + blocked records; e2e: entry order/metadata, cpuTimeMs excludes a 150ms handler wait, shim + import name resolution, handler errors (ok: false), failed runs keep the report, blocked entries at the limit, aborted-empty, prefix.run().
• typecheck, cargo fmt, eslint clean on changed files.