Accept every reasoning effort ReqLLM accepts, including max

[deepfates]

What changed

• Imp.Clients.ReqLLM reads its accepted :reasoning_effort values from ReqLLM's own generation schema (ReqLLM.Provider.Options.generation_schema(), the {:in, [...]} type of reasoning_effort) instead of keeping a second list, so Imp accepts max, which ReqLLM's list already had (since at least 1.18) and Imp's hand-written list lacked. The match fails the build if ReqLLM changes the option's shape.
• The saved-program loader (Imp.Saving) checks against the same list, so a saved "max" loads.
• A string effort is turned into its atom on its way into ReqLLM (the client's stored options, which feed the cache key, are unchanged). ReqLLM checks reasoning_effort against its atom list before most providers see it; only OpenAI, xAI and Meta convert the string form first. So in 0.6.0 a string effort, including every effort loaded from a saved program (always a string), failed every call on OpenRouter's default top-level wire, Anthropic, Google and Groq, with a NimbleOptions error. OpenRouter's nested wire, OpenAI and xAI were not affected.
• CHANGELOG: Unreleased entry.
• Dialyzer: the new code shifted the two .dialyzer_ignore.exs entries pinned by line in this file. The catch-all after ReqLLM.model/1's {:ok, _}/{:error, _} clauses in resolve_model/1 could never match (ReqLLM's contract); an unexpected result now raises a CaseClauseError that the function's own rescue turns into {:error, _}, so the clause and its ignore entry are removed. The response.provider_meta || %{} guard stays defensive (the struct does not enforce its map type); its entry is re-pinned to line 1902 with its existing reason. mix dialyzer.check passes locally.

Why

Some model residents are configured for the highest thinking level. ReqLLM accepts :max; Imp 0.6.0 refused it at build and per call.

Decision: accept every effort ReqLLM accepts, on every provider, and leave the mapping to ReqLLM. I read how ReqLLM 1.24 handles :max per provider: OpenRouter, Zenmux, Groq, xAI, Mistral and other pass-through providers send "max"; OpenAI's param profiles send "max"; Anthropic maps to "max" where the model supports it, otherwise "xhigh" or a fixed budget; Google maps to :high / 32768 tokens; LM Studio clamps to xhigh with a warning; Ollama and Fireworks send "max"; Moonshot uses :max. DeepSeek sends "max" as a string pass-through. None rejects it inside ReqLLM; whether a given model accepts "max" is the provider API's call, as it already is for xhigh.

Tests

• New: ReqLLMClientTest "reasoning_effort max reaches OpenRouter as "max" on either wire": a Req adapter (no network) captures the request body. "max" configured on the client arrives as top-level "reasoning_effort": "max"; :max on a call over a client with no effort does the same; openrouter_reasoning_wire: :nested sends "reasoning": {"effort": "max"}.
• New: ReqLLMClientTest "a string reasoning_effort reaches OpenRouter and Anthropic requests": the existing string "high" on OpenRouter's top-level wire arrives as "reasoning_effort": "high", and on Anthropic arrives as a thinking block.
• Extended: Imp.SavingReqLLMTransportTest "saved reasoning effort round-trips and remains narrowly allowlisted" round-trips :max through dump/load.
• Changed: the native-reasoning test ("ReqLLM registry capability drives the explicit typed reasoning contract") now sees :low rather than "low" reaching the ReqLLM module, which is the new conversion.

Falsification: with the old hard-coded list restored, the new OpenRouter test and the saving test fail (2 failures). With the string-to-atom conversion removed, the new OpenRouter test fails with ReqLLM's invalid value for :reasoning_effort ... got: "max", and the native-reasoning test fails. With lib/ taken from main, the string-effort test fails on its OpenRouter half and, with that half removed, on its Anthropic half, both with ReqLLM's got: "high" validation error. All restored: pass.

mix check after merging main (with #253) and the dialyzer fix: exit 0; 59 doctests, 9 properties, 3545 tests, 0 failures, 13 skipped (221 excluded).

Unsure / left alone

• bench/imp/benchmark_truth/rlm_manifest.ex keeps its own effort list. That manifest is shared with the DSPy runtime, so ReqLLM's list is not obviously its authority; left unchanged.
• No decisions.md line, to keep this PR off the files A model call's cost is the provider's reported charge; the catalog price is estimated_cost #253 touched. main (with A model call's cost is the provider's reported charge; the catalog price is estimated_cost #253) is merged in; CHANGELOG keeps both Unreleased entries.
• Imp.Clients.ReqLLM.reasoning_efforts/0 stays @doc false; making it public so Dwell can drop its own list is a follow-up.