Add fidelity / coverage columns to archive `turns` and `sessions`

[willwashburn]

Context

PR #78's archive schema deliberately omits coverage / fidelity columns:

Coverage / fidelity columns from the (still-open) coverage issue — the schema is deliberately additive so they slot in without a rebuild requirement.

Issue #40 calls out the requirement explicitly:

The archive should not blur:

• zero tokens
• missing tokens
• aggregate-only sources
• cost-only sources
• fully attributed per-turn sources

Issue #41 (coverage and fidelity metadata) and PR #76 are the upstream — they add fidelity flags to TurnRecord. The archive needs the equivalent columns so SQL queries can filter / group by them without re-deriving in memory.

Today, compare and summary cells render — instead of $0.00 / 0% for missing data; that decision currently lives in the analyze layer. Once the archive carries the flags, downstream consumers can express it as WHERE has_full_attribution = 1 rather than re-checking each row.

Proposal

1. Once Coverage and fidelity metadata: distinguish missing, zero, aggregate-only, and partial usage #41 / PR Add coverage and fidelity metadata to TurnRecord (#41) #76 lands the canonical fidelity flags on TurnRecord, add the columns to the archive's turns table additively (no rebuild required — ALTER TABLE ... ADD COLUMN with sensible defaults). Suggested:
   • attribution_fidelity TEXT — one of full, aggregate_only, cost_only, missing, zero.
   • tokens_present INTEGER — boolean: did the source give us per-turn token counts?
   • cost_present INTEGER — boolean: cost only?
2. Add session-level rollups to sessions:
   • min_fidelity TEXT — worst fidelity observed in the session.
   • has_full_attribution INTEGER.
3. Bump ARCHIVE_VERSION only if the additive ALTER strategy doesn't fit; otherwise, keep it stable so existing archives pick up the new columns on next build.
4. Update buildArchive() to read the fidelity flags from TurnRecord and persist them; update foldStamps() callers if any stamp source contributes to fidelity.

Tests:

• ALTER TABLE migration on an existing archive populates the new columns on next build.
• A mixed-fidelity fixture (Claude full + Codex aggregate-only) yields the right per-turn attribution_fidelity and the right session-level min_fidelity.
• compare / summary queries can filter by fidelity (WHERE attribution_fidelity = 'full') and produce the expected subset.

Acceptance criteria

• turns and sessions carry fidelity columns matching the upstream TurnRecord flags from Coverage and fidelity metadata: distinguish missing, zero, aggregate-only, and partial usage #41 / PR Add coverage and fidelity metadata to TurnRecord (#41) #76.
• Existing archives migrate forward (additive ALTER if possible; rebuild otherwise).
• burn archive status --json row-count summary surfaces a fidelity histogram.
• At least one consumer (likely burn compare) demonstrates filtering by fidelity through SQL rather than in-memory.

Out of scope

• Defining the upstream fidelity vocabulary (covered by Coverage and fidelity metadata: distinguish missing, zero, aggregate-only, and partial usage #41 / PR Add coverage and fidelity metadata to TurnRecord (#41) #76).
• Changing how — vs $0.00 is rendered in CLI output (the analyze layer already gates this).
• Coverage UI surfacing in burn archive status (text mode) beyond the JSON histogram.

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)
• Depends on Coverage and fidelity metadata: distinguish missing, zero, aggregate-only, and partial usage #41 / PR Add coverage and fidelity metadata to TurnRecord (#41) #76