feat(providers): add a dedicated DeepSeek provider module

[xianhongtao]

Summary

Adds a dedicated DeepSeek provider and refreshes WebBrain against the current
DeepSeek API (V4.1 line). The vendor-specific logic is moved out of the shared
OpenAI-compatible modules into two new files, so openai.js and
provider-compatibility.js stay protocol-generic.

Docs of record: https://api-docs.deepseek.com/zh-cn/ (模型 & 价格 / 思考模式 /
图像理解 / 上下文硬盘缓存 / 使用 Responses API / 限速与隔离).

Why

• The shipped default model id was the retired deepseek-v4-flash; the current
  id is deepseek-flash (DeepSeek-V4.1-Flash).
• deepseek-flash was not recognised by the capacity tables: it fell through to
  the 8K output ceiling and the generic 128K context window instead of 384K / 1M.
• Vision detection only knew deepseek-v4-flash-vision-exp, although the whole
  V4.1-Flash family is multimodal and deepseek-v4-pro is text-only.
• The card still carried pre-V4.1 prices and had no cache-hit rate, and
  DeepSeek's top-level prompt_cache_hit_tokens counter was ignored, so cached
  input was billed at the full input rate.

Architecture

Two new modules per build (src/chrome + src/firefox):

• providers/deepseek-config.js — the API contract: endpoints, shipped and
  retired model ids, model capacities, the reasoning-effort mapping, and the
  thinking / planner / vision request extras for both wire formats. It has no
  internal imports, so provider-compatibility.js can delegate to it without
  creating an import cycle.
• providers/deepseek.js — DeepSeekProvider extends OpenAICompatibleProvider,
  owning vision capability, stream_options.include_usage, and cross-turn
  reasoning_content replay (DeepSeek returns 400 when a tool-carrying
  follow-up omits it).

ProviderManager#_createProvider() dispatches the deepseek card — or any card
pointed at api.deepseek.com, or one that explicitly selects the deepseek
compatibility preset — to the dedicated class. Router-hosted DeepSeek models
(for example deepseek/deepseek-v4 on OpenRouter) stay on the generic provider.

Cleanup in the shared modules: openai.js loses its DeepSeek import, the
deepseek entry in _shouldRequestStreamUsage(), the DeepSeek branch of
_supportsReasoningContentReplay(), and the DeepSeek id in the vision regex; it
gains a _modelNameSniffedVision(model) hook that vendor subclasses override.
provider-compatibility.js loses mappedDeepSeekReasoningEffort() and the
hard-coded DeepSeek host/endpoint knowledge and now delegates. A source-guard
test fails the build if DeepSeek literals reappear in either shared module.

Behaviour changes

Model id	Context	Max output	Vision
deepseek-flash	1M	384K	yes
deepseek-v4-flash, deepseek-v4-flash-vision-exp (retired aliases, same model)	1M	384K	yes
deepseek-v4-pro (retired), deepseek-chat, deepseek-reasoner	64K	8K	no

Unknown DeepSeek ids deliberately fall into the conservative profile rather than
inheriting capacities we cannot verify.

• Thinking: top-level thinking object plus reasoning_effort; disabling
  thinking omits reasoning_effort entirely. The shared UI ladder maps
  minimal→low and medium/xhigh→high, matching DeepSeek's table.
• Wire format: Chat Completions stays the default. A new apiFormat
  selector in the Advanced panel opts into DeepSeek's Responses API, which then
  uses reasoning.effort (including none to disable thinking) and full
  text.format JSON Schema for the planner.
• Cost: off-peak list price (1 input / 0.02 cached input / 4 output CNY per
  1M, peak 2 / 0.04 / 8) converted at 1 USD = 7.1 CNY, with a new
  cacheReadCostPerMillionUsd. usage.prompt_cache_hit_tokens is now counted
  as cached input and priced at the cache-read rate, and the field is preserved
  in usage traces.
• Migration: cards that still hold the untouched shipped default (old model
  id, old base path and old prices, no API key, never saved) migrate to
  deepseek-flash and the new prices; saved or customised cards are never
  rewritten.
• Images: deepseek-flash is advertised as multimodal, so screenshots flow
  through the existing image_url path.

Also updated

• ui/settings.js: model suggestions drop the retired ids; the new apiFormat
  select is rendered only for the DeepSeek card, behind the Advanced panel.
• i18n: one new key (st.provider.field.api_format) added to all 23 locale
  files in both builds.
• Docs: docs/providers-and-models.md, docs/zh-CN/…, docs/fr/… — default
  model row, vision-detection table, a new DeepSeek section, and the contributor
  guidance now points at _modelNameSniffedVision() / vendor subclasses.

Verification

• node test/run.js → 2308 passed, 0 failed (includes new assertions for
  model capacities, the Responses wire shape, prompt-cache pricing, dispatch,
  migration, and the module-boundary guard).
• npm run test:provider-limits → passed.
• Manual wire checks: default requests go to /chat/completions with
  thinking + reasoning_effort; reasoningEffort: off clears
  reasoning_effort; apiFormat: responses sends only reasoning.effort;
  deepseek-v4-pro no longer advertises vision.

Notes / follow-ups

• user_id (DeepSeek's KVCache and scheduling isolation parameter) is not
  implemented; low concurrency limits (2500 for Flash) are expected to be fine
  for browser-agent traffic.
• The Anthropic-format endpoint (https://api.deepseek.com/anthropic) works by
  overriding the built-in anthropic card's base URL; documented, not a new
  provider card.
• Cost rates use off-peak prices, so peak-hour estimates are optimistic by 2×.

[vercel]

@xianhongtao is attempting to deploy a commit to the esokullu's projects Team on Vercel.

A member of the Team first needs to authorize it.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.