Relocate the parse-cancellation vocabulary out of llm_bridge.py (rename LLMBridgeCancelled family)

[danielaskdd]

Summary

PR #3605 promoted the LLMBridgeCancelled exception family and the normalize_cancel_events / first_cancellation helpers into the repo-wide "a blocking parse wait was cancelled" vocabulary (its docstring now says so explicitly), and gave them their second and third non-LLM consumers. The vocabulary itself, however, still lives in lightrag/parser/llm_bridge.py — an LLM-named module — and the class names still say "LLMBridge". This is follow-up naming/placement work, deliberately kept out of the security PR (it touches pipeline.py, which that PR's diff otherwise never does).

Why it matters

lightrag/pipeline.py keys its cancelled-vs-FAILED document classification on LLMBridgePipelineCancelled (the except near the parse dispatch, ~line 4441). Any blocking parse path that wants cancellation treatment must raise from this family — raising anything else silently records the document as FAILED instead of cancelled.

A contributor adding cancellation to the next blocking path (e.g. the MinerU/Docling result downloads tracked in #3610, or docx work) has no reason to look inside llm_bridge.py for the exception contract of a non-LLM code path. PR #3605 mitigated discoverability with a docstring; relocation is the durable fix.

Proposed change

1. Create lightrag/parser/cancellation.py holding:
   • the exception family (renamed to match its actual role, e.g. ParseCancelled / ParsePipelineCancelled / ParseShutdown);
   • normalize_cancel_events and first_cancellation.
2. Update all importers directly — lightrag/pipeline.py, lightrag/parser/native_base.py, lightrag/parser/llm_bridge.py, lightrag/parser/markdown/parser.py, and tests. These are internal APIs (underscore-free but unexported/undocumented), and per this repo's convention internal refactors do not leave deprecation shims or re-exports behind — no llm_bridge aliases.
3. SyncLLMBridge itself stays in llm_bridge.py and imports the vocabulary from the new module.

Sequencing

Blocked on PR #3605 merging — that PR touches llm_bridge.py and markdown/parser.py, so doing this move in parallel guarantees conflicts. Branch off main after the merge.

Context

• Raised during the review of fix(parser): bound native markdown image downloads per document (GHSA-25c3-j78v-83qx) #3605; the PR author flagged the same naming overreach in the PR body ("For the reviewer" section) and proposed handling it separately.
• Related follow-up: Bound external adapter result downloads (MinerU/Docling): unbounded body buffering, uncapped zip extraction, no wall-clock deadline #3610 (bounding the MinerU/Docling result downloads) is precisely the kind of new blocking path whose implementer needs to find this contract.