fix(core): notify reloads immediately and debounce at bursty sources - #46843

Why

State treated its two change paths differently: a registration or disposal notified immediately, while reload() waited out a 500 ms trailing debounce first. The debounce assumed every reload came from a bursty file watcher. Counting the callers on v2, that is true for three of them and wrong for the rest:

• Explicit operations. Mcp.add and Mcp.remove reload so their reconcile creates or drops the server entry, and ACP prompts right after mcp.add, so they must wait for reconcile — but they were idling for the debounce first. Every mcp.add sat for 500 ms before the connection even started.
• Multi-domain pairs. Seven sites reload Integration then Catalog after auth changes and Ollama/LM Studio/vLLM/Models.dev discovery; two serial debounces meant the catalog update reached the TUI about a second after the models were already readable.
• Config-driven reloads. config.updated is already debounced 100 ms in Config, so those paid a second debounce for nothing.

The three genuinely bursty consumers (instruction scanner, skill scanner, MCP ToolsChanged) were relying on a side effect: the consumer being blocked in reload() for 500 ms let a PubSub.sliding(1) or runForEachArray coalesce the burst behind it. Measured with a 3- and 6-event editor save burst against the instruction scanner, that produced two notifications per burst, not one.

What Changes

• reload() notifies immediately and awaits the notification, exactly like transform and dispose. state.ts loses reloadDebounce, requestedAt, pending, and the detached debounce loop (−32 lines); changed is one code path.
• The three bursty consumers debounce at the source with Stream.debounce("100 millis"), matching Config, agent.ts, and command.ts. Measured: the same save bursts now produce one notification.
• Debouncing a Stream forks its upstream pull into a child fiber, so the PubSub subscription happens one fiber hop later than a plain runForEach and an event published in that hop is lost. All three sites therefore subscribe first (PubSub.subscribe, scoped) and debounce the already-open subscription (Stream.fromSubscription). tool/mcp.ts bridges the bus into its own sliding PubSub for the same reason; the bus subscription itself stays eager as before.
• State.batch is unchanged and remains the way to coalesce several changes into one notification; it no longer doubles as the escape hatch from a debounce.
• Tests: the debounce-specific State tests are replaced by immediate-notification equivalents; twelve fork → TestClock.adjust(500) → join dances around reload() collapse to yield* reload(); test/lib/clock.ts no longer mentions a State debounce.

Verification

# packages/core
bun typecheck
bun run test                                    # 3,993 pass, 40 skip, 0 fail
# packages/cli
bun run test test/acp                           # 96 pass, 0 fail
# packages/server
bun run ../core/script/test.ts                  # 50 pass, 3 skip, 0 fail

Burst measurement (temporary harness, not committed): instruction scanner with 3 and 6 watcher events 2 ms apart — before: 2 notifications per burst; immediate reload without a source debounce: 3 and 6; with the source debounce: 1 and 1.