Tighten finetuning HTTP API reference

[EthanReid]

Summary

This trims the finetuning HTTP API reference back toward Vik's tighter version while keeping the real API additions we still want to document.

A lot of the previous edits had drifted into a more verbose, overly detailed style. This PR rolls much of that back so the page reads like an HTTP reference again instead of a trainer/SDK handoff doc.

What changed

• Removed verbose or non-contract-heavy sections that did not need to be exposed in the HTTP docs
• Removed mention of checkpoint download from the HTTP reference
• Removed the extra metrics/train-log sections that were added later
• Restored the finetuned inference section closer to Vik's original structure
• Simplified the rollouts section and kept only the rollout metadata note that clients need for RL round-tripping
• Updated POST /train_step so the RL example clearly shows both request and rollouts
• Added a compact Targets subsection for SFT with a small table and concrete snippets for query, point, and detect

Accuracy fixes

• Fixed GET /finetunes/:finetuneId to show the actual wrapped response shape: { "finetune": { ... } }
• Fixed the point/detect settings docs to match the current API
• Removed the incorrect max_objects: 50 default
• Restored max_tokens for point/detect in the request examples where that had been documented incorrectly
• Kept max_objects as detect-only in the settings table
• Updated wording like slower -> slower training
• Removed misplaced notes like the query SFT note from the ground-truth section

Verification

• Cross-checked the doc against tuna/src/tuna/rl/api_types.py
• Cross-checked the control-plane routes/response shapes against md-api
• Rendered locally with the Docusaurus dev server and reviewed the updated page

[Copilot]

Pull request overview

This PR tightens the finetuning HTTP API reference to be more “contract-focused” while updating several endpoint schemas/examples to reflect the current API (notably /finetunes/:finetuneId, /rollouts, and /train_step).

Changes:

• Corrects GET /finetunes/:finetuneId to show the wrapped response shape ({ "finetune": { ... } }).
• Updates /rollouts docs/examples (e.g., num_rollouts, spatial_refs, max_tokens defaults, rollout metadata note).
• Expands /train_step documentation to cover both RL and SFT modes, adds a “Targets” subsection, and documents a non-empty response shape with metrics.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.