fix(s08): snip_compact respects max_messages and handles tool-result boundaries

[qiugu]

Background
snip_compact is the L1 layer of the s08 context-compaction pipeline. It drops the middle of the message list (replacing it with a placeholder) once the count exceeds max_messages. The original implementation had three issues.

Problems
Off-by-one (result exceeds the limit): The kept region keep_head(3) + keep_tail(max-3) = max, plus the inserted placeholder, yields a result of length max + 1 — so max_messages was never actually honored. When len(messages) == max + 1, it "removed 1, added 1 placeholder" and effectively did nothing.
Out-of-bounds for small max_messages: keep_tail = max_messages - 3 becomes <= 0 when max_messages <= 3, producing an empty/negative tail window, and the early-return then handed back the still-over-limit list.
Tail only handled a single tool_result: The original tail_start -= 1 left orphan tool_results at the tail start when a single tool call produced multiple results, and was asymmetric with the while loop used on the head side.
Fix
Reserve one slot for the placeholder: keep_tail = max_messages - keep_head - 1, so the compacted result never exceeds max_messages.
Add a max_messages < 2 guard to avoid empty/negative tail windows.
The tail boundary now walks back with while over consecutive tool_result messages, keeping multi-result tool calls intact (symmetric with the head side).
keep_head is capped with min(3, max_messages - 1) to stay safe for small max.
Verification
All 6 cases in tests/test_compaction_tool_pairs.py pass.
Additional checks: off-by-one (7 → ≤6), large scale (100 → ≤50), tail with multiple tool_results leaves no orphans, multi-result tool-call boundary, and max_messages=1 is a no-op — all as expected.
Scope
Only snip_compact in s08_context_compact/code.py changed. The contract is unchanged: head/tail are preserved, tool_use/tool_result pairs are protected, and the [snipped N messages] placeholder is used.