feat(billing): settle at the tier the request's input length selects - #417

Summary

Models can now declare optional [[models.tiers]], and settlement bills a request at the price step its input length selects.

Step 1 (496a723) landed the field and its validation but left settlement untouched, so every request was billed at the first step. This step:

• carries the steps into the database — models.tiers_json (schema v16 → v17), written by db::seed_models from config and read back by dao::get_model_price in the same query it already ran;
• selects the step at settle time in billing::select_tier, on this request's input tokens (cache-miss + cache-hit prompt tokens), falling back to the model's own single price when there are no steps or the input is below the first from_tokens;
• keeps peak-hour prices orthogonal: peak_* is a model-level trio and applies on top of the selected step.

from_tokens is an inclusive lower bound ([from_tokens, next)). Vendors that publish an open interval are expressed by adding 1 in the data: MiniMax's "input >512k" → 512001, xAI's and Google's "prompt exceeds 200k tokens" → 200001.

Tier data is filled only where the vendor publishes a number (checked against the pricing pages on 2026-10-16/17):

Rows deliberately left single-priced: OpenAI splits short/long context but does not publish the boundary token count; Anthropic's pricing page has no long-context step; qwen3.7-flash publishes input/output per step but no cache-hit price, so writing a step would drop its cache-hit price to 0 (under-charging). Each is noted as pending, with the reason, in the config header.

The admin model editor now clears tiers_json on write: it is a config-owned field, and a hand-edited base price would otherwise leave stale steps charging long prompts. Manual edits fall back to the single price, and the next start re-seeds the row from config.

Cache-write price (cache_write_input_per_m) stays declaration-only in this step, and the config comment now says so: no upstream response path yields a cache-write token count today, so wiring it into billing needs an extra usage parse.

Related Issue

References rant 2026-10-16T14:08:14.003832+08:00 (tiered price and cache-write price), step 2 — settlement. Step 1 (496a723) landed the expression and validation.

Changes

• Feature/fix described
• Config/data-structure changes synced to the example file (config/config.example.toml) and the database roster (docs/architecture.md)

Tests

• cargo test all green — 579 passed / 0 failed (8 new)
• cargo fmt --check passes
• New unit tests added
• cargo clippy --all-targets -- -D warnings and a non-test cargo build with -D warnings are clean

New legs: select_tier boundaries at from_tokens−1 / from_tokens / from_tokens+1 (and past the last step); a settle-level test that drives the real path (seed_models → dao → select_tier → ledger) and reads the recorded points back across the 200k boundary; a leg fixing the selection basis to cache-miss + cache-hit; and tiers_json round-trip plus malformed-JSON fallback legs. Both boundary legs were shown to bite by mutating the implementation (dropping the cache-hit term; making the lower bound exclusive) and watching them go red.

Checklist

• Branch naming follows the convention (feat/)
• Commit message uses Conventional Commits
• Single responsibility, minimal change