Migrate @relayburn/mcp tool handlers to archive queries

[willwashburn]

Context

The MCP server's two tool handlers — currentBlock and sessionCost — both call queryAll() from @relayburn/ledger (packages/mcp/src/tools/current-block.ts:6,47, packages/mcp/src/tools/session-cost.ts:3,28). Every in-session self-query from a running agent therefore re-walks the entire ledger and re-folds stamps. The MCP path is explicitly the place where latency matters most — the agent is paying per-token to wait for the answer.

PR #78 explicitly defers this rewire:

Rewiring burn summary / compare / plans / @relayburn/mcp to read from the archive (each command is a self-contained migration that keeps the in-memory fallback intact).

Issue #40 calls this out as a primary motivation for the archive:

@relayburn/mcp (#26) wants low-latency self-query from inside a running agent.

The MCP tools accept a queryTurns dependency injection in their Deps types already (current-block.ts:35-40, session-cost.ts:21-25), so the rewire surface is small.

Proposal

1. Add a queryTurnsFromArchive(query) helper in @relayburn/ledger that prepares a parameterized SQL SELECT against the archive's turns table and returns the same EnrichedTurn[] shape callers already consume. The archive already carries workflow_id / agent_id / persona / tier plus the JSON enrichment blob, so the result shape can be reconstructed without folding.
2. Default currentBlock's queryTurns and sessionCost's queryTurns to the archive-backed helper.
3. Run buildArchive() (incremental tail apply) at MCP server startup so the first tool call doesn't pay the cold cost. Optionally re-run on a low-frequency interval, or on-demand before each call if the ledger has grown — measure first.
4. Fallback path: if the archive open or query throws, log via onLog and fall through to queryAll(). The MCP server must never fail a tool call because the archive is unavailable.

Tests:

• Existing MCP unit tests pass with the archive-backed helper injected via Deps.
• New end-to-end test: spin up the MCP server against a fixture archive + ledger, call both tools, assert results match the ledger-backed equivalents.
• Fallback path: archive missing/corrupt -> tools still succeed via queryAll.
• Cold-start path: first MCP call after a new ledger tail applies the tail incrementally before responding.

Acceptance criteria

• burn__currentBlock and burn__sessionCost issue SQL against archive.sqlite on the hot path, not full ledger walks.
• Tool responses are equivalent (within float-rounding tolerance for cost) to the pre-migration implementation.
• Server startup includes an incremental archive build; subsequent tool calls do not re-walk the ledger.
• Archive failure paths fall back transparently to queryAll().
• Latency: tool-call P50 on a 100k-turn ledger is at least 10x faster than the current implementation.

Out of scope

• New MCP tools or new tool inputs.
• Schema additions to the archive.
• The CLI summary/compare/plans rewires.

Refs

• Spun out of Derived analytics archive foundation: archive.sqlite + burn archive (#40) #78 (Derived analytics archive: materialize the ledger into a local queryable store #40 follow-up)
• Related: @relayburn/mcp: MCP server exposing ledger data for in-session self-query #26