feat: refresh model catalog and DeepSeek vision

[woai3c]

Goal and acceptance criteria

• Goal: Refresh the curated provider model catalog and make DeepSeek V4.1 Flash use native vision without borrowed vision, OCR, or local-image fallback.
• Acceptance criteria:
  • Current curated model IDs, defaults, context/output limits, and reasoning tiers are documented and tested.
  • DeepSeek V4.1 Flash sends user and tool-result images natively; DeepSeek V4 Pro remains text-only.
  • Direct attachments do not trigger redundant local re-reading, and image de-duplication state is invalidated whenever history loses its binary content.
  • Explicit DeepSeek reasoning levels use native reasoning_effort without the xhigh to max compatibility warning.
  • pnpm 10.28.2 is selected consistently for local and CI workflows.

Scope and plan

• Scope and non-goals: Provider catalog metadata, provider request mapping, image attachment handling, CLI attachment presentation, tests, documentation, and pnpm selection. This PR does not release or merge the change.
• Approach: Centralize current model profiles in the provider catalog, use native DeepSeek image transport and reasoning options, prevent redundant image reads with correctly invalidated delivery state, and derive workflow pnpm versions from packageManager.
• Affected packages/interfaces/state: @x-code-cli/core providers and agent media flow, CLI transcript labels, root documentation/configuration, and CI workflows.
• Risks, assumptions, and Unknown questions: Provider availability and limits can change after this revision. DeepSeek paths received live runtime verification; other provider entries are protected by local contract tests and documented source research.
• Plan path: Full path. The repository maintainer approved implementation and local acceptance in the originating task conversation.

Evidence

• Local Fast Verification: PASSED on macOS with Node v22.23.2 and corepack pnpm@10.28.2 run ci; typecheck, lint check, format check, build, unit, fault, package, and PTY suites passed. 203 test files passed, 1 skipped; 2022 tests passed, 27 skipped.
• Targeted Runtime Verification: PASSED with DeepSeek V4.1 Flash. A real image was analyzed natively in one turn without readFile, shell image processing, OCR, or another vision model. A subsequent hello request completed in one turn and the fresh debug log contained no reasoning-tier mapping warning.
• Full E2E: NOT EXECUTED. The full multi-provider real-LLM suite requires additional provider credentials and cost; targeted live DeepSeek verification covers the changed runtime path, while local CI and PR CI provide the fallback for the remaining behavior.
• Cross-platform evidence or exclusions: Local verification ran on macOS. Platform-neutral unit/fault/package/PTY coverage passed locally; GitHub PR CI is the next trigger for the configured OS matrix.
• Independent pre-acceptance Agent review: PASSED for commit 6dab7a8 by a fresh empty-context reviewer. One P2 cache-invalidation finding was fixed, regression-tested, and closed on re-review with no remaining findings.
• Human Local Acceptance: PASSED. Repository maintainer stated 代码可以提交了 in the originating task conversation.
• PR Human Technical Review: Pending on this Ready PR.
• Platform AI review: NOT CONFIGURED.
• Material evidence artifacts: Local CI output, targeted DeepSeek runtime debug trace, regression tests for native image delivery/reasoning mapping/cache invalidation, and the independent review result.

Deviations and knowledge

• Plan deviations or unresolved risks: The initial independent review found stale image delivery state after provider rejection or transcript compression. The implementation now clears that state whenever binary history is removed or rewritten, with regression coverage. No unresolved correctness findings remain.
• Durable knowledge: Observed DeepSeek native image and reasoning behavior is captured in provider metadata, request-shape tests, attachment-flow tests, and README documentation. Provider catalog claims remain time-sensitive and should be refreshed against official provider documentation when availability changes.

Readiness

• Non-trivial plan review is recorded, or the bounded fast path is justified.
• Applicable local and runtime outcomes are passed or explicitly accounted for.
• Applicable independent pre-acceptance Agent review is complete, or its unavailable capability and fallback are recorded.
• Human Local Acceptance is recorded or an established policy is cited.
• Risks, unknowns, E2E status, and evidence exclusions are visible above.