fix(perps-controller): order 0: Reduce only order would increase position

[abretonc7s]

Explanation

Fixes the remaining TAT-3252 failure paths. A stale position or half-up size rounding could submit a reduce-only order larger than the current position. #9719 fixed single-position close only.

Core now uses one position-source rule:

• symbol action: current DEX slice, otherwise HTTP;
• batch action: complete current WebSocket snapshot, otherwise complete REST snapshot;
• never merge REST and WebSocket positions.

This covers TP/SL, single and batch close, reduce-only edits, margin updates, and HIP-3 margin sizing. No public API changes.

Proof

• Reproduced first: on pre-fix source, regressions submit an oversized TP/SL size and read stale position data, reproducing the Jira failure path.
• Proved after: the same regressions pass. A live HyperLiquid testnet recipe passed 47/47 nodes on d4e527eb85: updateMargin succeeded, a 100% partial TP rested at no more than the position size, the position stayed open, and cleanup closed it.
• The full Perps suite, TypeScript, ESLint, formatting, and changelog checks pass.

References

• Fixes TAT-3252
• Also fixes the close half of TAT-3873
• Follows fix(perps): prevent reduce-only close orders from increasing positions #9719, which covered only single-position close

Checklist

• Tests, JSDoc, and changelog updated
• Breaking changes introduced

---

Note

High Risk
Changes core trading paths (close, batch close, TP/SL, margin) and WebSocket position cache semantics; mistakes could submit wrong reduce-only sizes or skip live positions.

Overview
Hardens HyperLiquid reduce-only flows (close, batch close, TP/SL, margin updates, HIP-3 margin sizing) so order size and side always come from one fresh source, not caller snapshots or a frozen aggregate cache.

Position freshness: WebSocket per-DEX slices are stamped with a connection epoch that advances on socket close/reconnect; stale clearinghouseState from a replaced client is ignored. Symbol-scoped actions use the current DEX slice or a single HTTP read; closeAll/empty symbols require a complete all-DEX WebSocket snapshot or a strict REST fan-out (PROVIDER_NOT_AVAILABLE if any DEX is missing). REST and WebSocket positions are never mixed. closePositions loads only DEXes for the requested symbols and can return per-symbol unavailable results while still closing others.

Sizing: Partial TP/SL, whole-position TP/SL formatting, and reduce-only editOrder sizes are floored to the venue grid (with early ORDER_TPSL_SIZE_INVALID / ORDER_SIZE_POSITIVE); non-reduce edits keep half-up rounding.

API/docs: Optional batch-level error on ClosePositionsResult; optional position on close/TP/SL is documented as diagnostic only.

^{Reviewed by Cursor Bugbot for commit 2ab8e68. Bugbot is set up for automated code reviews on this repo. Configure here.}

[geositta]

The reduce-only sizing and position freshness work is well tested, and the current CI suite passes. I am requesting changes for two exit-path regressions that affect production behavior:

1. Closing selected positions now depends on loading position data for every enabled market group.
2. A delayed position update from a replaced WebSocket connection can be accepted as current after reconnect.

Please scope selected position closes to the market groups represented by the requested symbols, and reject callbacks from replaced subscription clients. The batch level error documentation should also be aligned with the result shapes returned at runtime.

I found no additional TypeScript or numeric precision concerns, and this PR does not change React code.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d30dfbe. Configure here.}