Migrate `burn summary` from ledger walk to archive query

[willwashburn]

Context

burn summary today calls queryAll() from @relayburn/ledger, which streams the entire ledger.jsonl from disk and re-folds every stamp in memory on every invocation (packages/cli/src/commands/summary.ts:35, packages/ledger/src/reader.ts). PR #78 landed the derived archive at ~/.relayburn/archive.sqlite with materialized turns rows that already have stamp enrichment columns (workflow_id, agent_id, persona, tier) and indexes on ts, model, activity, project_key, and workflow_id — exactly the dimensions summary filters and aggregates on.

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).

Proposal

Replace the queryAll() call in runSummary with an archive-backed query:

1. Call buildArchive() after ingestAll() so the archive is current before we read.
2. Add a small summarizeFromArchive(query) helper to @relayburn/ledger that issues a single SQL SELECT ... FROM turns WHERE ... GROUP BY model (or the per-row projection summary needs) using the existing indexes.
3. Keep the queryAll() path behind a fallback flag (--no-archive or env RELAYBURN_ARCHIVE=0) so users can validate parity and we have an escape hatch if the archive is missing/corrupt.
4. Subagent-tree and by-subagent-type modes still need per-turn rows; for those, either pull from the archive or fall back to queryAll. The archive already materializes subagent_id / parent_subagent_id / subagent_type so the tree mode should work natively.

Tests:

• A summary fixture that asserts the archive-backed and ledger-backed code paths produce the same output (text + --json) for a non-trivial mixed-stamp ledger.
• A test that the archive is auto-built if missing before the first summary invocation.
• A test that RELAYBURN_ARCHIVE=0 (or the equivalent flag) routes through queryAll unchanged.

Acceptance criteria

• burn summary no longer scans ledger.jsonl line-by-line on the hot path; it issues SQL against archive.sqlite.
• Output is byte-identical to the pre-migration implementation for the parity fixture (text and --json).
• Subagent-tree (--subagent-tree) and --by-subagent-type modes work against the archive (or transparently fall back).
• Fallback path (--no-archive flag or env) preserves the old behavior.
• Performance: on a ledger with >=100k turns, archive-backed burn summary --since 7d is at least 5x faster than the current ledger walk on the same machine. (Numbers in the PR description.)

Out of scope

• Schema additions to the archive (this is a read-side migration).
• Coverage / fidelity columns (separate follow-up).
• Rewiring compare, plans, or MCP tools (separate follow-ups).

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)