Move stateful message handling into AgentMessageHandler (step 4b)

[shellicar]

Completes the AgentMessageHandler extraction by moving the stateful cases in.

What moves in:

• Helper functions fmtBytes, primaryArg, formatRefSummary, formatToolSummary (were module-level in runAgent.ts)
• Private fields: #lastUsage, #usageBeforeTools
• message_usage case: delta token calculation, marginal cost annotation via appendToLastSealed
• tool_approval_request case: snapshots usageBeforeTools, fires #toolApprovalRequest
• tool_error case
• Private #toolApprovalRequest async method (permission check, approval prompt, respond callback)

Constructor change: adds opts: AgentMessageHandlerOptions — model, cacheTtl, cwd, store, tools, respond callback. The respond callback abstracts port.postMessage so the handler doesn't hold a reference to port.

runAgent.ts after this: 89 lines. Sets up tools, calls agent.runAgent(), constructs respond + handler, wires port.on('message', handler.handle). No message logic lives there anymore.

Restores: the message_compaction context-usage annotation ([compacted at X/Y (Z%)]) that was temporarily omitted in 4a because it needed lastUsage.

Tests: 27 total (16 unchanged + 11 new). New tests cover tool_error, message_usage with/without tool batch, delta annotation values, reset-timing sequence (two batches compute independently), and message_compaction annotation presence.

[bananabot9000]

Stateful extraction complete, regression fixed 🍌

State properly encapsulated: #lastUsage and #usageBeforeTools are private fields on the handler. Three stateful cases moved in: tool_approval_request, tool_error, message_usage. Four helper functions relocated. Constructor takes typed AgentMessageHandlerOptions with respond callback abstracting the message port.

4a regression fixed: message_compaction now reads this.#lastUsage for the [compacted at X/Y (Z%)] annotation. Both present and absent paths tested.

runAgent.ts is now 89 lines. No message logic remains. Just tool setup, agent.runAgent(), handler instantiation, and port.on('message', handler.handle). A thin wiring function.

27 tests (16 existing + 11 new). The delta reset-timing tests are the highlight -- proving two independent tool batches compute deltas independently, snapshot resets between batches. This was the risky part and it's well covered.

Observations (not blocking):

• void this.#toolApprovalRequest(msg) -- fire-and-forget async, matches original pattern, void makes intent explicit. Good.
• removePendingTool duplicated in try + catch -- could be finally, but explicit is fine.
• tool_approval_request not directly tested (async + mock complexity). Acceptable scope boundary for this step.
• Unicode escapes replacing literals (\u00b7, \u2190, \u2013) -- functionally identical, likely editor normalization.

Ship it 🚢🍌