feat(server): encode DeepSeek V4 reasoning effort

[pepuscz]

Summary

Add the official DeepSeek V4 Flash low, high, and max reasoning-effort
encodings to the native renderer.

• low adds no model-facing prefix
• high and max use the exact official prompt prefixes
• accepts reasoning.effort, top-level reasoning_effort, and
  chat_template_kwargs.reasoning_effort
• maps API-compatible medium and xhigh to DeepSeek Flash high
• keeps Lucebox's hyphenated x-high extension as a separate budget tier,
  using the model-facing max encoding
• uses DeepSeek's high model-facing default when thinking is enabled without
  an effort; disabling thinking suppresses all effort prefixes and budgets
• keeps bare chat_template_kwargs.thinking and enable_thinking booleans
  renderer-only; only thinking.type: enabled or an explicit effort opts into
  the force-close budget envelope and finish_details
• advertises the official DeepSeek effort set through /props
• uses the shared rendering helper for normal requests, FlowKV rerenders, and
  agent-turn-cache canonicalization

Official references:

• https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/blob/main/encoding/README.md
• https://api-docs.deepseek.com/api/create-chat-completion/
• https://api-docs.deepseek.com/api/create-response/

Rebase

Rebased onto main at c994209, including PRs #640 and #647. The server
rendering-helper conflict was resolved by preserving add_generation_prompt
and passing the normalized effort through the shared helper.

Validation

• fail-before/pass-after regressions for renderer-only DeepSeek and Qwen
  controls versus budget-envelope opt-in
• request-policy tests for all effort aliases, all three request spellings,
  precedence, default effort, renderer overrides, and explicit token budgets
• exact full-prefix renderer tests for low, high, max, thinking-disabled,
  and no-generation-prompt rerenders
• full HIP test_server_unit: 423/423 passing on gfx1201
• git diff --check
• full GitHub CI matrix green on final head 10514b5: Windows, hosted
  CUDA/C++ server tests, R9700, Strix Halo, RTX 3090, DGX GB10, workspace, and
  speed-profile workflows

[cubic-dev-ai]

No issues found across 6 files

_{Re-trigger cubic}

[pepuscz]

Rebased onto current main (e79f1b6), preserving the merged DeepSeek named-tool-call coverage while resolving the documentation and unit-test conflicts. The rebased commit is 1924d7c. Validation on the gfx1151 ROCm host: dflash_server builds successfully; test_server_unit passes 370/370; test_feature_gate passes 1/1; and the focused reasoning-prefix, tool-capability, and reasoning-capability tests each pass independently. GitHub Actions for the new head are currently awaiting maintainer approval.

[davide221]

Rebased and requalified against current main (c994209, including #640 and #647).

Final head: 1616e3d

Key corrections made during the rebase/review:

• preserved the shared current-main renderer for normal requests, FlowKV rerenders, and agent-turn-cache canonicalization
• corrected DeepSeek Flash aliases: medium/xhigh use model-facing high; Lucebox's hyphenated x-high extension uses model-facing max while retaining its own budget tier
• defaulted explicitly enabled DeepSeek thinking without an effort to high
• ensured disabling thinking removes the effort prefix and request budget caps
• added exact full-prefix rendering coverage plus request precedence, alias, toggle, and budget tests

Validation is green on the rebased head: full GitHub CI (Windows, hosted Linux/CUDA and C++ server tests, Strix Halo gfx1151, R9700 gfx1201, RTX 3090, and DGX GB10), focused tests against the R9700-built test_server_unit binary, and git diff --check.

The PR is now conflict-free and ready for merge. It has not been merged.

[davide221]

Correction to my earlier ready-for-merge note: a second contract audit found one real blocker in the previous head. Bare chat_template_kwargs.thinking / enable_thinking booleans were incorrectly setting thinking_opt_in. That could activate force-close budgets and finish_details even though these compatibility booleans are renderer controls, and the behavior also leaked into Qwen requests.

Fixed in 21fddbc:

• renderer-only booleans now change prompt rendering without activating the budget envelope
• thinking.type: enabled and explicit reasoning effort remain the two budget opt-ins
• a final disable clears prior effort opt-in and per-request caps
• DeepSeek renderer-only enable still uses the model-facing high default, but without an implicit force-close cap
• the canonical budget spec and DS4 documentation now state this distinction

The regression tests were run against both versions: they fail on 1616e3d at the incorrect cap/opt-in assertions and pass on 21fddbc. The rebuilt gfx1201 HIP unit binary passes 423/423 tests, including focused DeepSeek request, prefix-rendering, and Qwen compatibility coverage. git diff --check is clean.

The PR remains one commit directly on current main (c994209) and is conflict-free. Please wait for the new GitHub CI matrix before merging; the old green checks belong to the superseded head.

[davide221]

Final audit follow-up: head 10514b5 changes the no-explicit-effort request log label from low to default. The previous label was misleading for renderer-enabled Qwen requests; this is diagnostics-only and does not change request behavior. The rebuilt gfx1201 HIP unit binary still passes 423/423. A fresh GitHub matrix is now running for this final head.

[davide221]

Final status for 10514b5: all required validation is green.

• local gfx1201 HIP unit binary: 423/423
• GitHub hosted CUDA/C++ build and server tests: pass
• Windows MSVC + CUDA library/server build and smoke test: pass
• R9700, Strix Halo, RTX 3090, and DGX GB10 jobs: pass
• workspace/import and speed-profile workflows: pass
• PR is one commit on c994209, conflict-free, and git diff --check is clean

With the renderer-only budget-opt-in regression fixed and covered, I now consider this head ready to merge. It has not been merged.