fix(token-usage): decide the long-context tier once per priced entry

price_entry computed is_long_context for the recorded column while the
per-shape cost formulas each recomputed it to select rates — consistent
only by coincidence of calling the same function twice. The decision is
now made once and threaded into the cost, so the recorded flag (which
the wire's long-context splits are built from) can never diverge from
the tier the cost actually billed at.

Cost/db tests also move from claude-sonnet-4-20250514 to a model id the
embedded snapshot actually carries: models.dev dropped first-party
claude-sonnet-4, and the tests had been passing through the family
fallback's (formerly leaked, now neutralized) tier data.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>