Fix DeepSeek output defaults and HF route context budgets - #957

Summary

Avoid reserving the model's maximum possible output on every native DeepSeek request, without moving compaction earlier or reducing model capability metadata.

• Native DeepSeek: omit default max_output_tokens, preserve explicit overrides.
• HF Together V4-Flash-0731 and Novita V4.1-Flash: omit default output limits.
• HF DeepInfra V4-Flash-0731: default to 131,072 output tokens (not a hard cap); omission still reserves 384,000 upstream.
• HF Scaleway V4-Flash-0731: use its actual 262,144-token context for display/compaction, retaining the 32,768 output cap.
• Document behavior and test defaults, model-query/per-call overrides, and compaction boundaries.

Motivation

A fast-agent 0.10.24 trace terminated with a context-limit 400: 655,743 input + 393,216 reserved output exceeded 1,048,576 by 383 tokens. Input alone still fit, and full-window 85% compaction had not triggered.

Validation

• 148 focused native DeepSeek and HF router-profile tests passed.
• uv run scripts/format.py --check
• uv run scripts/lint.py
• uv run scripts/typecheck.py
• git diff --check

Live probes (reasoning=max, tiny requested reply)

Native DeepSeek Responses with output limit omitted completed at 656,071 and 890,071 input tokens.

HF chat probes near 890K input:

• Together V4-0731: explicit model-wide allowance failed with context overflow; omission passed.
• Novita V4.1: explicit allowance returned 400; omission passed (the 400 did not expose its precise cause).
• DeepInfra V4-0731: explicit model-wide allowance and omission both failed with a 384,000-token reservation; explicit 131,072 passed.
• Scaleway V4-0731: reported 262,144 context; follow-up passed at 222,920 input with 32,768 output allowance.
• Baseten (both models), Fireworks V4.1, and Novita V4-0731 passed both variants. Fireworks V4-0731 passed omission; its explicit-limit probe timed out (504), so inconclusive.

Initial approximately 7 MB HF requests hit HTTP 413; shorter padding reduced payload size and allowed the token-budget probes above. Probes establish acceptance at these boundaries, not maximum generation behavior. No credentials or trace artifacts are included.

Contributor question

You're given a calfskin wallet for your birthday. How would you feel about using it?

I'd appreciate the gesture, but prefer a non-animal alternative.