fix(pricing): deterministic price tiebreak; admit a zero-cost reserve with a zero hold (#98, #104)

[JumpTechCode]

#98 — non-deterministic current-price resolution

resolve_price ordered by published_at DESC LIMIT 1 with no tiebreak. price_book.published_at defaults to func.now() (the transaction timestamp, constant within a transaction), so a bulk seed/migration publishes several versions at the same instant and the "current" pick is arbitrary — and can differ run to run if physical order shifts. The resolved version is stamped on the reservation and drives commit reconciliation, so an ambiguous pick silently shifts the cost basis.

Fix: ORDER BY published_at DESC, version DESC, making "latest published" a total order.

#104 — a genuinely free model can't transit the gate (decision: admit with a zero hold)

reserve raised NonPositiveEstimate (422) whenever the worst-case estimate rounded to 0. Because ceil_micro keeps any positive worst case at ≥ 1 micro, estimate == 0 happens only for a zero-token request or a zero-priced (free) model — so a free model always 4xx'd on reserve.

Fix (per the resolved question): admit a zero estimate with a zero-micro hold so free and paid models route through the same guard. The reserve walks every node at amount 0 — which succeeds on any node with non-negative headroom (ensure_period seeds the row, the guard is headroom >= 0) — and commit later reconciles to 0. estimate_micro never returns negative, so estimate >= 0 always holds.

Dead-error cleanup

NonPositiveEstimate was raised only by reserve, so admitting zero makes it an error the server can never emit — the same vestige pattern the review flags in #96 (unused status column) and #100 (dead 402). Removed the domain type and its api/errors + SDK mappings. The SDK's error_for already falls a 422-with-unmapped-code back to InvalidRequest (what non_positive_estimate mapped to), so client behavior is unchanged.

Tests

• Integration: two versions sharing a published_at resolve deterministically to the higher version; the existing latest-published test unchanged.
• Unit: a zero-token reserve is now admitted with a zero hold (walks nodes at amount 0, persists estimated_micro=0, caches the response) instead of raising; NonPositiveEstimate cases dropped from the error-mapping tests.

Local gate: ruff, mypy --strict, import-linter, full unit suite (489), price-book + reserve integration — all green.

Closes #98, #104.