feat(scheduler): per-adapter KV prefix-cache namespace + max_loras batch cap

[qywu]

Summary

Two scheduler-level changes that make the tokenspeed scheduler LoRA-aware. Rebased onto current main from the previously-closed #268 (extracted from the broader LoRA serving PR #83).

---

1 — Per-adapter KV prefix-cache namespace isolation

Problem: Without isolation, tokens cached under adapter A could be incorrectly reused by adapter B or the base model.

Fix: Thread a lora_id through KVPrefixCache::Insert/Match and the radix tree so each adapter gets an isolated namespace off a per-adapter root. lora_id == kLoraNone (0) → shared base-model root (unchanged behavior when LoRA is off). The id is plumbed from Request::LoraId() through every FSM event that publishes or matches KV pages: SchedulePrefillEvent, ScheduleDecodeEvent, FinishEvent, ExtendResultEvent, and the InsertHybridCache helper.

2 — max_loras batch cap

SchedulerConfig::max_loras (0 = LoRA disabled). newForwardOperation defers requests that would push the unique-adapter count in a batch over the cap to the next step, guaranteeing prepare_loras() never sees n_unique > max_loras. Exposed through the Python scheduler config and nanobind binding.

---

Rebase / merge notes

This reconciles the original change with several things main grew since #268:

• FSM event constructors gained a conditional KvCacheCoordinator* (flat-kvcache) parameter upstream; lora_id is appended after it so both features compose in flat and non-flat builds.
• InsertHybridCache was promoted to a header-declared helper with a prefix_pages_override parameter upstream; lora_id is appended as the trailing defaulted argument.
• ExtendResultEvent is new since the original PR and also publishes to the KV prefix cache — wired it through with the request's lora_id to preserve the per-adapter isolation guarantee (would otherwise be a silent cross-adapter cache hole).

Testing

• All changed scheduler translation units type-check against current main headers (via compile_commands.json, -fsyntax-only).
• clang-format and pre-commit run --all-files are clean.
• ⚠️ A full C++ build / gtest run was not possible in this environment (missing tokenspeed-spdlog at CMake-configure time). Recommend running the scheduler C++ test suite in CI before merge.

🤖 Generated with Claude Code