blackbox: cpa proxy turns → lane cpa + served vendor/provider, priced estimated (stacked on #1392)

[ang-fleet-workers]

Card t_d59c7936. Stacked on #1392 (native kimi-oauth + K3 snapshot + Kimi in _infer_vendor_from_model). This branch contains #1392's commits; merge #1392 first, then this rebases to one commit (7fc5851).

Problem

A cpa/kimi-k3 turn (CLIProxyAPI, which serves Kimi, Codex and Grok behind one key) was recorded as provider=cpa, lane_family='other', with no vendor and no cost (billing_mode=unknown).

Change

• agent/usage_pricing.py
  • NOTIONAL_PROXY_PROVIDERS = {cpa, kimi-code, kimi-cpa, cliproxyapi}, also reached as custom:<name>.
  • resolve_billing_route dispatches a proxy turn on the SERVED model's vendor. It uses the one existing map, _infer_vendor_from_model: moonshotai goes to the kimi-oauth route (K3 snapshot), openai to openai-codex (OpenRouter notional), xai to xai-oauth. All three are estimated. An unknown vendor stays unpriced.
  • codex-* joins the openai vendor prefix.
  • attribute_route(provider, model) returns {vendor, vendor_label, served_provider}. cpa/kimi-k3 gives moonshotai, Kimi (Moonshot), kimi. An unknown model gives vendor-unknown, never other.
• plugins/blackbox/store.py
  • lane_family never returns other. Proxy lanes return cpa, a custom:<name> lane is unwrapped, anything else returns its own provider name, and an empty provider returns unknown.
  • The backfill now calls lane_family() and no longer keeps a SQL copy of the map. It also re-derives rows already stored as other.
  • Guarded additive columns: turns.lane_family/vendor/served_provider and turn_api_calls.vendor/served_provider, written at insert.
• plugins/blackbox/last_turn.py: proxy turns get a • Route: vendor Kimi (Moonshot) · provider kimi · lane cpa line. Other lanes are unchanged.

Not in scope: the subs.ace and tokens.ace renderers (scripts/subs_data in hermes-home) belong to the sibling Kimi usage card, which consumes attribute_route and the new columns.

Verified

• tests/plugins/blackbox/test_cpa_lane_attribution.py has 52 tests. They cover kimi, gpt, codex and grok through every alias and custom:cpa, pricing parity with the native lanes ($0.00465 for 1k/100/500 on K3), the vendor-unknown negative control, every providers.list_providers() name and alias mapping to a named family, the stored columns, the backfill retiring other, and the last-turn line. All 52 pass. With store.py and last_turn.py reverted, 41 of them fail.
• Neighbour tests pass: blackbox, usage_pricing, vendor fallback and kimi-oauth, 456 in total. test_codex_auto_review_stays_unknown still passes.
• Not verified yet: the live cpa/kimi-k3 turn on tokens.ace. That waits for the batched restart.

Override: none by design — the 4 refusal lines are #1392's kimi-oauth device-flow terminal errors (access_denied / expired_token in hermes_cli/auth.py), carried here only because this PR is stacked on #1392; a user who denied the browser login must log in again, no operator bypass applies. This PR adds no refusal of its own.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[ang-fleet-lander]

🤖 merged-by: apollo · lane: kanban-merge-pass · gate: ADVISORY (FleetReview not green for 4071e0e): fleetreview-advisory-20260927-standing.md · why: t_d59c7936: Blackbox: CLIProxyAPI (cpa) turns attributed to vendor/provider Kimi (etc.) with; Argus off card review (Ace 13:08), CI green

[ang-prism]

FleetReview

Review: post-merge · head fbfcf03a309b · duration 13m 31s
Profile: light (merit: default light: lines 297<800, files 4<1000000, hunks 15<1000000, no hot path) · policy: below-size-and-path-gates
Roster: B-state → gpt-6-sol (openai), C-assert-xhigh → claude-code-opus-5-5 (anthropic), G → gpt-6-sol (openai), L6 → gpt-6-sol (openai)

Post-merge review (fleetreview:post-merge override): this reviewed the merge commit against its first parent — the bytes that already shipped. It is not a pre-merge gate pass.

Reviewed with 2 of 3 model families — xai unavailable.

profile: light (rule: default light: lines 297<800, files 4<1000000, hunks 15<1000000, no hot path) · round 0 · members: B-state, L6, C-assert-xhigh, G · families: anthropic,openai

Confidence: 3/5

Findings

• P1 agent/usage_pricing.py:2093 — Unsupported proxy vendors still receive estimated prices · agreed: C-assert-xhigh,G (anthropic, openai)

---

FleetReview provenance · models: B=gpt-6-sol, C=claude-code-opus-5-5, D=grok-4.6, F=gpt-6-sol · cost: $2.55 · duration: 13m 29s · rounds: 2 · files examined: 4