Bug: CitationRegistry global state causes cross-request citation contamination

Bug Description

CitationRegistry stores its state in a class-level dictionary (_instances) shared across all requests. init_citation_registry() calls CitationRegistry.reset(), which wipes this global dict for every concurrent session. When two requests run the init_citation_registry → assign_citation_ids_stateful pipeline at the same time, one request's reset destroys the other's in-progress state, producing corrupted or swapped citation IDs in responses.

Location

servers/custom/src/custom.py, ~line 405:

class CitationRegistry:
    _instances: Dict[int, Dict[str, Any]] = {}  # class-level, shared across all requests

    @classmethod
    def reset(cls):
        cls._instances = {}                      # wipes state for ALL concurrent sessions

~line 435:

@app.tool(output="q_ls->q_ls")
def init_citation_registry(q_ls: List[str]) -> Dict[str, Any]:
    CitationRegistry.reset()                     # global reset triggered per request
    return {"q_ls": q_ls}

Reproduction

1. Send two concurrent requests that both invoke init_citation_registry followed by assign_citation_ids_stateful.
2. Request A calls reset() while Request B is mid-way through assign_citation_ids_stateful.
3. Request B's accumulated citations are wiped; it returns citation IDs starting from 1 for documents it had already assigned higher IDs.

Impact

Users receive incorrect citation numbers in answers, causing documents to be cited under wrong IDs. In multi-tenant deployments this also constitutes a cross-session information leak (one user's citation state can be reset by another user's request).

Suggested Fix

Scope registry state per request using a unique session/request ID rather than global class state:

def init_citation_registry(q_ls: List[str], request_id: str) -> Dict[str, Any]:
    CitationRegistry._instances[request_id] = {}
    return {"q_ls": q_ls, "request_id": request_id}

Or pass a fresh CitationRegistry instance through the pipeline context instead of using class-level storage.

Found via automated codebase analysis. Happy to submit a PR if this is confirmed.

Start in servers/custom/src/custom.py around CitationRegistry, init_citation_registry, and assign_citation_ids_stateful. Trace how registry state moves through the concurrent request pipeline and compare the proposed request-scoped or instance-based approaches. Done means concurrent requests retain isolated citation state and return correct, non-swapped citation IDs without cross-session resets.