fix(pricing): price gpt-6.1-sol codex usage rows

[mikehasa]

Problem

Codex is reporting gpt-6.1-sol, and nothing in agentacct priced it. The live ledger showed it plainly:

Client  Model        Rows  Why
codex   gpt-6.1-sol    11  no local price row covers it

Two rows of the price stack both missed it:

• the builtin table stops at gpt-5.5 (gpt-5.6* was deliberately left to the snapshot, which is why codex/gpt-5.6-sol has rows too — those stay excluded for a different reason: codex reported no input/output split for them);
• the local LiteLLM snapshot was fetched before the model appeared upstream, so the codex -> openai provider alias had nothing to resolve to.

Stored rows do not heal on their own: a scan only reprices what a catalog row now covers, so the 11 rows would have stayed cost_confidence: unknown until a snapshot refresh happened to pick the model up.

Change

The builtin table now carries gpt-6.1-sol at its published rates, on both routes a codex row can report — the client's own ("codex", ...) key and OpenAI's ("openai", ...) key:

	per 1M tokens
input	$2.00
output	$10.00
cache read	$0.10
cache write (5m)	$2.50

Source: LiteLLM's model_prices_and_context_window.json (gpt-6.1-sol, litellm_provider: openai), the same upstream table the snapshot is built from.

• No multiplier. The 2.5x ccusage fast-pricing convention stays scoped to gpt-5.5, so the codex and openai routes price identically rather than one silently carrying a convention. Cost basis is pricing_table / estimated_from_tokens, never provider-billed.
• Both keys, like gpt-5.5. Exact-key lookup answers for the codex route instead of depending on the snapshot's TTL refresh to alias it onto OpenAI's row; if a snapshot does carry the model, the builtin row still answers for its exact codex key and the dollars agree.
• Coverage stays per model. A sibling id no row carries (gpt-6.1-sol-pro) stays cost-unknown rather than borrowing this row's price.

Tests

• tests/test_pricing_catalog.py (+2): both routes resolve to the builtin row with the published rates, no multiplier, cache read/write columns priced through estimate_model_cost_breakdown_usd; sibling/uncovered ids stay unpriced; and a snapshot that does carry the model leaves the price and provenance unchanged.
• tests/test_stored_usage_repricing.py (+1): the stored-row reprice path prices a codex gpt-6.1-sol row from the builtin table with no snapshot at all, and produces the exact row a fresh import does (dollars + pricing_source provenance), leaving the stored row untouched while planning.
• pytest -q tests/test_pricing_catalog.py tests/test_stored_usage_repricing.py → 53 passed.
• Full suite: pytest -q → 4894 passed, 7 failed. All 7 are in tests/test_onboard_global.py and are pre-existing and unrelated — they fail identically on unmodified origin/main (an environment string drift: the test asserts the word "yourself" appears in onboarding copy this machine does not produce).

Verification against a real ledger

Run on a copy of a real store, with the pricing snapshot deleted so nothing but the builtin table is in play, and auto-refresh pinned off (no network):

usage import-local --store-dir <copy> --client codex --refresh --estimate-costs

→ 11/11 gpt-6.1-sol rows priced, cost_confidence: estimated_from_tokens, cost_basis: pricing_table, pricing_source: agent_sentinel_builtin, $6.528786 total. The copied store's 11 stored rows are the only thing written; the live store was never touched. Spot-checked one row by hand: 5,859 in × $2/1M + 400 out × $10/1M + 125,440 cache read × $0.10/1M = $0.028262, matching the stored value exactly.

Docs / safety

• CHANGELOG.md under Unreleased → Fixed.
• Preserves local-first behavior (no network on the builtin path); no keys or secrets; no API calls were run — the prices come from the published upstream table.