Fix reasoning_effort tier mapping and context-gate misreading

[Flor1an-B]

Follows up on #635 (@kernelzeroday, @zom-2018) and #653 (@rell666) — this is the patch discussed there, posted as a comment first (#635 (comment)) and opened as a PR per @darkbasic's question.

Two independent bugs, same two functions

1. DS4_THINK_HIGH never rendered a prefix at all. Only DS4_THINK_MAX did, using text that is actually the official "high" tier prefix per deepseek-ai/DeepSeek-V4-Flash-0731's encoding/README.md (verified verbatim against the primary source). Every reasoning_effort other than "max" silently rendered as official "low". The real "max" prefix ("Beyond maximum — exhaustive, relentless, and uncompromising...", note the em dash U+2014) was absent from the codebase entirely — unreachable via the API regardless of setting. Fixed in both independent prompt-rendering paths that had this bug: ds4.c's chat_push_think_prefix() (CLI/engine) and ds4_server.c's render_deepseek_chat_prompt_text() (server).

2. ds4_think_mode_for_context() downgraded "max"→"high" below a 384K context threshold, citing DeepSeek's guidance as an input-context requirement. The model card's actual text: "For the high and max reasoning effort levels, we recommend a maximum output length of 384K tokens." That's an output-budget guideline, not a context-size precondition. Turned into a documented no-op rather than removed outright, since four ds4_server.c call sites use it and ds4_think_max_min_context() may still be useful for client-side output-length advice.

Testing

• Adds test_render_think_high_prompt_prefix() — there was previously zero test coverage of the "high" tier ever rendering anything.
• Fixes the now-incorrect context-gate assertion in test_reasoning_effort_mapping().
• Adds a mutual-exclusivity assertion to the existing "max" prompt test.
• make ds4_test passes clean against the full existing suite plus these changes.
• Smoke-tested live against a real 0731 quant: reasoning_effort: high and : max now produce visibly distinct prompts; : max no longer collapses to the "high" text below 384K context.

Requirements

• I have read and agree with the contributing guidelines.
• AI usage disclosure: YES — diagnosed and patched with Claude Code, verified against the primary source (DeepSeek's own model card) and the project's existing test suite.

---

Update: a real DS4_THINK_LOW tier (issue #660)

Found while looking for other reasoning_effort-related community reports to help with: issue #660 ("New 0731 iq2 thinks forever!") describes reasoning_effort: "low" producing 9k tokens of runaway reasoning.

Root cause, again verified against DeepSeek-V4's own encoding spec (encoding/README.md, "Reasoning Effort" section): it defines exactly three levels — "low" (the default, prompt prefix "none"), "high", and "max". DS4's parse_reasoning_effort_name() collapsed "low" (and "medium"/"minimal") into the same DS4_THINK_HIGH enum value as an explicit "high" request — the code even had a comment acknowledging this: "DS4 only exposes HIGH and MAX above zero, so minimal collapses to the smallest non-zero level (HIGH)." So a client asking for "low" got the most elaborate non-max prompt DS4 has, instead of the model's own default behavior with no extra prompting — a very plausible direct cause of runaway reasoning on tasks that don't need it.

Added DS4_THINK_LOW as a genuine fourth ds4_think_mode value ("thinking enabled, no prompt prefix", distinct from DS4_THINK_NONE which disables thinking entirely) and wired it through every switch/branch over the enum. The default reasoning_effort when a client sends none at all is untouched (still high); this only fixes what an explicit "low"/"minimal" request does.

Tested the same way as the rest of this PR: full suite passes, new regression test (test_render_think_low_prompt_prefix) plus updated test_reasoning_effort_mapping, and a live A/B on the production Flash quant confirming reasoning_effort: "low" and "high" now produce genuinely different reasoning_content (they were byte-identical prompts before this fix).