Add missing Wafer serverless models

[hgezim]

Adds the five models listed at https://docs.wafer.ai/serverless/setup#models that were missing from providers/wafer.ai: GLM-5.3, GLM-5.3-Flash, Kimi-K3, DeepSeek-V4-Flash-0731-Fast, and Qwen3.5-397B-A17B.

Pricing, context, and reasoning-effort tiers taken from the live catalog (https://pass.wafer.ai/v1/models), which the docs cite as authoritative. Qwen3.5-397B-A17B has no published Wafer pricing yet, so its cost block is intentionally omitted with a header comment explaining why.

[github-actions]

Action items

• [medium] [possible mistake] providers/wafer.ai/models/GLM-5.3.toml:6 - Check: Relay reasoning_options must match the lab/same-surface baseline unless this host truly exposes a different control. Why: First-party providers/zhipuai/models/glm-5.3.toml and peers (TokenGo, OrcaRouter, Vancine, Hyper) use always-on effort low|high|max with no off control; this file adds none. That can mislead callers into disabling thinking on a model the lab documents as non-disableable. Action: Confirm from Wafer’s model metadata that reasoning_effort=none actually turns thinking off for GLM-5.3; if not, drop none and use ["low", "high", "max"] (same for evidence if the catalog only advertises a generic enum).
• [medium] [possible mistake] providers/wafer.ai/models/GLM-5.3-Flash.toml:6 - Check: Same host-vs-lab reasoning baseline for GLM-5.3-Flash. Why: Lab/peers are effort low|high|max only (thinking not disableable); this entry adds none. Action: Verify Wafer really disables Flash thinking on none; otherwise remove none and align to ["low", "high", "max"].
• [medium] [possible mistake] providers/wafer.ai/models/Kimi-K3.toml:7 - Check: Baseline effort set for Kimi K3 on this host. Why: Lab Moonshot and common openai-compat peers expose graded low (with high/max), often plus off via toggle or none. This file omits low and only lists none|high|max, relying on a collapse claim. Action: Verify Wafer metadata that low is not a distinct level (collapses); if low is accepted as its own tier, add "low" to values (e.g. ["none", "low", "high", "max"]).

[github-actions]

Action items

• [medium] [possible mistake] providers/wafer.ai/models/GLM-5.3.toml:8 - Check: Reasoning options must match this host’s real controls, with extra levels beyond lab/peers backed by host-specific evidence. Why: Lab GLM-5.3 is always-on (low/high/max only; thinking cannot be disabled). Almost all peers match that. This file adds none as off via Wafer effort. If none is only accepted/ignored or maps to a minimum effort, callers will think reasoning is off when it is not. Action: Verify on Wafer that reasoning_effort=none actually disables GLM-5.3 thinking (not just that the catalog enum lists it). If it cannot disable, use ["low", "high", "max"] only; keep none only if off is a real, distinct wire effect.
• [medium] [possible mistake] providers/wafer.ai/models/GLM-5.3-Flash.toml:9 - Check: Same host-vs-lab reasoning baseline for GLM-5.3-Flash. Why: Lab Flash is also always-on low/high/max. This entry again adds none as off. Same risk of advertising a disable control the model/serve does not honor. Action: Same verification as GLM-5.3; drop none unless Wafer truly turns Flash thinking off for that value.
• [medium] [possible mistake] providers/wafer.ai/models/Kimi-K3.toml:8 - Check: Relay effort set should follow lab/same-surface peers unless this host collapses tiers. Why: Moonshot K3 is toggle + low/high/max. This file uses only none/high/max and omits low, citing collapse. If Wafer still exposes a distinct low (or a separate thinking.type on/off), the catalog will understate controls. Action: Confirm Wafer’s per-model metadata/live behavior for Kimi-K3: if low is distinct, include it; if off is a separate toggle field rather than effort=none, use toggle + graded effort without none in the effort list (with a leading wire comment).
• [low] [possible mistake] providers/wafer.ai/models/Qwen3.5-397B-A17B.toml:9 - Check: Interleaved reasoning field when this host generally returns one. Why: providers/wafer.ai/provider.toml and the other new Wafer reasoners set [interleaved] field = "reasoning_content". This Qwen entry enables a thinking toggle but omits interleaved, so clients may miss side-channel reasoning if Wafer emits it. Action: Confirm Wafer’s Qwen response shape; if reasoning lands in reasoning_content, add the same [interleaved] block as the sibling models.