output_limit::spill_aware_collect and helpers are dead code after GH #133 item 2

[tobert]

Context

GH #133 item 2 (fixed in a companion PR) removed dispatch.rs::BackendDispatcher::try_external's only call site for crate::output_limit::spill_aware_collect — the twin's non-pipe capture branch now uses BoundedStream + drain_to_stream (matching production's kernel.rs::try_execute_external) instead.

After that fix, the following are unreferenced anywhere in the codebase except their own doc comments and unit tests:

• output_limit::spill_aware_collect (pub async fn, #[cfg(feature = "subprocess")])
• output_limit::collect_stderr (private helper)
• output_limit::collect_stdout_with_spill (private helper)
• output_limit::handle_overflow (private helper)
• output_limit::drain_in_memory (private helper)
• output_limit::extend_ring (private helper)
• output_limit::stream_to_spill (private helper)

Roughly ~200 lines of implementation plus their dedicated unit tests (output_limit.rs tests around the duplex-stream-based collect_stdout_with_spill/extend_ring cases).

Why this is a separate issue, not part of #133

spill_aware_collect is declared pub async fn inside pub mod output_limit — it's reachable as kaish_kernel::output_limit::spill_aware_collect by embedders (kaibo, kaijutsu), even though it isn't re-exported at the crate root. Removing it is a breaking embedder-API change (per CLAUDE.md's changelog policy), which deserves its own deliberate PR + CHANGELOG entry + review, not a side effect of a test-double behavioral-sync fix. Flagged during kaibo's review of the #133 item 2 PR.

Suggested follow-up

• Confirm no embedder actually depends on spill_aware_collect (check kaibo/kaijutsu).
• Remove spill_aware_collect and its now-unreferenced private helpers, plus their dedicated tests.
• Add a **BREAKING:**-prefixed CHANGELOG entry under Removed.