fix: support DeepSeek V4+ model names in normalization

[zons-zhaozhy]

Problem

DeepSeek V4 models (deepseek-v4-pro, deepseek-v4-flash) are unusable. The /model command reports:

Model `deepseek-v4-pro` was not found in this provider's model listing.

Root Cause

_normalize_for_deepseek() assumed DeepSeek only has two model IDs (deepseek-chat and deepseek-reasoner). All other names were forced to deepseek-chat. When a user requests deepseek-v4-pro, it gets mangled to deepseek-chat, which is then rejected by the live API model validator since the API only returns deepseek-v4-pro and deepseek-v4-flash.

Additionally, DeepSeek returns "Insufficient Balance" with HTTP 400 (not the documented 402) when credits are exhausted. This was misclassified as a model-not-found error rather than a billing error, masking the real problem.

Changes

hermes_cli/model_normalize.py — Refactor DeepSeek normalization

• Before: Whitelist-based (deepseek-chat or deepseek-reasoner only)
• After: Prefix-based passthrough — any deepseek-* name passes through unchanged
• Alias resolution for short forms: v3 → deepseek-chat, r1 → deepseek-reasoner
• Reasoner keyword detection and backward-compatible fallback preserved
• Future DeepSeek models (V5, V6...) work without code changes

agent/error_classifier.py — Add billing pattern

• Added "insufficient balance" to _BILLING_PATTERNS
• Ensures DeepSeek credit exhaustion is classified as billing error, not model-not-found

hermes_cli/models.py — Update static catalog

• Added deepseek-v4-pro and deepseek-v4-flash to the DeepSeek provider model list

tests/hermes_cli/test_model_normalize.py — 14 new test cases

• Passthrough: 6 models including future-proof deepseek-v5-ultra
• Short alias resolution: v3, r1
• Reasoner keyword detection: thinker, cot-model, reasoning-v2
• Unknown fallback to deepseek-chat
• Vendor prefix stripping: deepseek/deepseek-v4-pro → deepseek-v4-pro

All 69 tests pass (55 existing + 14 new).

Testing

$ pytest tests/hermes_cli/test_model_normalize.py -v
============================== 69 passed in 0.43s ==============================

[alt-glitch]

Likely duplicate of #15123 (merged) — same fix: pass DeepSeek V-series IDs through instead of folding to deepseek-chat. Also overlaps with #14946 (closed) and #14911 (closed). The error_classifier billing pattern addition may be the only net-new content.

[teknium1]

Thanks for the contribution! This is an automated hermes-sweeper review.

The core fix described in this PR — DeepSeek V-series passthrough normalization and catalog entries — was already landed on main in commit 4ac731c84 (fix(model-normalize): pass DeepSeek V-series IDs through instead of folding to deepseek-chat), which also covered the _DEEPSEEK_V_SERIES_RE regex and deepseek-v4-pro/deepseek-v4-flash catalog entries in hermes_cli/models.py.

• hermes_cli/model_normalize.py line 144: _DEEPSEEK_V_SERIES_RE regex already present
• hermes_cli/models.py lines 268–269: deepseek-v4-pro and deepseek-v4-flash already in catalog
• Member @alt-glitch noted this overlaps with fix(model-normalize): pass DeepSeek V-series IDs through instead of folding to deepseek-chat #15123 (merged)

Two items worth noting for follow-up:

1. insufficient balance billing pattern — this string is not yet in _BILLING_PATTERNS (agent/error_classifier.py line 90). A focused one-line PR or issue to add it would be welcome.
2. Progressive compression feature (commit 48400a7) — this is a substantial unrelated feature bundled into this PR. If you'd like to contribute it, please open a separate PR focused on that work.