fix(spend_logs): billing_status='partial' when cancel dispatched to upstream

[songkuan-zheng]

Tier classification

• D — Universal mechanism + company opinion in litellm/ core

Tried upstream first?

• No — same dogfood window as PR feat(cancel): billing accuracy + orthogonal delivery/billing status taxonomy #78's Tier-D candidate (cancel taxonomy). Will be folded into the upstream issue once v1.87.0-internal.N has one release of production data confirming the corrected mapping. Tracked in UPSTREAM_PR_QUEUE.md.

Summary

Refines the _derive_delivery_billing_status rules introduced in PR #78. The old derivation mislabeled two cancel scenarios as (delivery=none, billing=none) even though the SpendLogs row carried a positive spend from compute_prompt_only_cost in the failure-hook path.

What was wrong

Two cancel paths produced rows where metadata.billing_status="none" contradicted the row's own spend column:

Scenario	phase	bytes	spend	Old billing	Real
Streaming cancel before first chunk	streaming_partial	0	compute_prompt_only_cost > 0	none ❌	partial
Non-stream shield gave up (no_completion)	during_upstream	0	compute_prompt_only_cost > 0	none ❌	partial

In both cases the proxy had already dispatched the prompt to upstream (we were inside the streaming generator or the non-stream upstream wait). Upstream received the prompt and started consuming tokens. The failure-hook then bills the prompt-only baseline. The derived billing_status must reflect that, otherwise dashboards aggregating by billing_status will under-report cancel revenue.

What's fixed

Reorder the derivation so phase=='before_upstream' is the ONLY path that returns (none, none). Everything else inside the cancel branch falls through to (none, partial) when no positive-delivery or shield-success signal matches.

The updated semantic mapping in the docstring:

| scenario                                       | status   | delivery | billing |
| streaming cancel WITH chunks                   | success  | partial  | partial |
| streaming cancel BEFORE first chunk            | success  | none     | partial |  ← new
| non-stream shield success                      | success  | none     | full    |
| non-stream shield_timeout                      | success  | none     | partial |
| non-stream cancel during upstream wait         | success  | none     | partial |  ← new
| cancel BEFORE any dispatch (before_upstream)   | success  | none     | none    |  ← only legit none/none cancel
| real failure                                   | failure  | none     | none    |

Verification

• ✅ 217/217 pass across tests/test_litellm/proxy/spend_tracking/ + tests/test_litellm/litellm_core_utils/test_cancel_*.py
• The two affected tests were renamed + their assertions inverted:
  • test_zero_chunk_cancel_yields_none_none → test_before_upstream_cancel_yields_none_none (unchanged behaviour, only path that still returns none/none)
  • test_streaming_cancel_with_zero_bytes_yields_none_none → test_streaming_cancel_with_zero_bytes_yields_none_partial
• Added test_cancel_during_upstream_no_completion_yields_none_partial.

Conflict resolutions

N/A — single commit, branch linear from ship/v1.87.0 HEAD d1dbe6b004 (post-#80 merge).

Pre-submission

• Tests in tests/test_litellm/
• make test-unit equivalent passes (217/217)
• PR scope: 1 commit, surgical fix to the rule shipped in feat(cancel): billing accuracy + orthogonal delivery/billing status taxonomy #78