feat(cost): add pricing metadata and resolve prefix shadowing

[kesslerio]

Summary

This PR adds fallback model pricing metadata for several new flagship and cloud models and refines the fallback resolver logic to prevent prefix shadowing.

Changes

• Model Pricing Expansion: Added fallback pricing metadata for new models including gpt-5.5, gpt-5.4-mini, kimi-k2.6:cloud, deepseek-v4-flash:cloud, glm-5.1:cloud, grok-4.3, qwen3-coder-next, and minimax-m3:cloud.
• Model ID Fallback Resolution: Updated resolvePricingFromFallbackMap to extract the trailing model ID component (e.g. gpt-5.5 from openai/gpt-5.5) for direct matching. Additionally, fallback keys are now sorted by specificity (length descending) before suffix matching to prevent shorter keys (like gpt-5) from shadowing longer prefix-matching keys (like gpt-5.5).
• gpt-5.4-mini Price Alignment: Aligned the output pricing rate of gpt-5.4-mini to $4.50/M to match the repository's pinned openclaw provider metadata.

Verification

• Added test cases in tests/server/cost-utils.test.js validating correct pricing calculations for provider-qualified model names openai/gpt-5.5 and openai/gpt-5.4-mini.
• Ran the full test suite (npx vitest run) to confirm clean execution.

---

[chrysb]

Thanks for splitting this out. I think the prefix-shadowing fix is the right direction, but the new fallback prices should be aligned with the OpenClaw plugin manifests before merge. I checked the bundled OpenClaw manifests in node_modules/openclaw/dist/extensions and found several mismatches:\n\n- openai gpt-5.5 / gpt-5.4-mini have cacheWrite: 0 in the manifest, but this PR sets cacheWrite to the input rate.\n- moonshot kimi-k2.6 is 0.95 input / 4 output / 0.16 cacheRead / 0 cacheWrite in the manifest, but this PR uses 0.8 / 3 / 0.1 / 0.8.\n- deepseek-v4-flash cacheRead is 0.028 in the manifest, but this PR uses 0.0028.\n- glm-5.1 is 1.2 input / 4 output / 0.24 cacheRead / 0 cacheWrite in the manifest, but this PR uses 1.4 / 4.4 / 0.26 / 1.4.\n\nThis is not a runtime-breaking issue, but it is a cost-dashboard correctness issue: when the fallback table is used, especially for provider-qualified or :cloud refs, we would show confidently wrong totals. I would fix the table to mirror the OpenClaw manifests, and ideally add a test for one of the :cloud fallback paths so future updates do not drift.