Fix KeyError when spot token index is missing from tokens list

[RWCS-LTD]

Summary

Info.__init__ raises KeyError (or IndexError on pre-dict-refactor versions) and every client crashes at startup when the HL API briefly returns a universe pair whose token index is not present in spot_meta["tokens"].

Reproducer

Observed in production on 2026-05-29: HL added spot pair @367 referencing token index 479 while the tokens list held 464 entries. Every Info() caller crashed for ~54 minutes until HL repaired the response shape.

Quick repro against live API at the time of the incident:

```python
import json, urllib.request
body = json.dumps({"type": "spotMeta"}).encode()
r = urllib.request.urlopen(urllib.request.Request("https://api.hyperliquid.xyz/info", data=body, headers={"Content-Type": "application/json"}), timeout=10)
spot_meta = json.loads(r.read())
n = len(spot_meta["tokens"])
bad = [(u["name"], u["tokens"]) for u in spot_meta["universe"] if max(u["tokens"]) >= n]
print(f"Pairs with out-of-range tokens: {bad}")

Pairs with out-of-range tokens: [('@367', [479, 0])]

```

Fix

Skip universe pairs whose base or quote is not in token_by_index:

```python
if base not in token_by_index or quote not in token_by_index:
continue
```

This lets Info() initialize for all valid pairs and keeps clients alive through any future universe/tokens drift on the API side. New valid pairs come back into scope automatically as soon as their tokens land in the list.

Test

Added `tests/test_info_spot_tokens.py` with a synthetic `spot_meta` that includes one valid pair and one pair pointing at a non-existent token index. The test asserts `Info()` no longer raises and that the valid pair is still registered.

Verified the test fails on `master` without the patch (`KeyError: 5`) and passes with it.

Test plan

• `pytest tests/test_info_spot_tokens.py` passes with the patch.
• Same test fails on `master` without the patch.
• Hand-tested against the live API after HL repaired the response shape — `Info()` initializes correctly and lists 463 of 464 spot tokens (the malformed pair is silently skipped, as intended, until HL adds the missing token).

[Choppaaahh]

Production data point in support of this fix.

We have run exactly this skip in place — if base not in token_by_index or quote not in token_by_index: continue inside the spot_meta["universe"] loop — since June 2026 on a long-lived market recorder. We patched it after a live spotMeta response (2026-06-09) contained a pair whose token index (479) exceeded the tokens list length (465), which took Info() down at construction. With the skip, the recorder has run continuously since; on clean days (e.g. 2026-08-08: 484 tokens / 324 pairs / 0 bad indices) the guard never fires, so there is no behavioural change when the data is well-formed.

Two observations that may help the maintainers decide:

• The failure mode is intermittent on the live API, so a unit test alone can't reproduce it — the value is in fail-safe construction, not in reproducing the bad payload.
• The skip is strictly safer than raising: a malformed spot pair is unusable anyway, and raising takes down perps-only consumers that never touch spot.

Happy to contribute an additional regression test (an index-out-of-range pair alongside valid ones) if that would help this land.

---

AI-assisted.