[Refactor]: Extract attention-sink trimming rule into named behavior

[AashishH15]

Problem

Attention-sink trimming is critical behavior. It’s currently implemented but should be reinforced by naming to prevent accidental removal.

Evidence

assembler.rs (line 69) explains the trimming rule in comments. The rule is important, but the function can carry more of the explanation through names.

Scope & constraints

No response

Proposed approach

Rename/extract helpers like:

• trim_context_tail_preserving_attention_sink
• fits_budget
• preserve_attention_sink_head

Acceptance criteria

• No behavior change.
• Policy is encoded in names (self-documenting).

Test plan

• cargo test llm::assembler.