fix: reduce false positive compaction detection

[KorenKrita]

Problem

Compaction detection (the C count shown in context info like 200k C35) was producing massive false positives. The original logic detected compaction whenever context size (inp + cr) dropped >30% between turns:

// Before: any context drop > 30% counts as compaction
if prev_context > 0 && result.last_context_tokens < prev_context * 7 / 10 {
    compaction_count += 1;
}

In agentic loops, cache_read fluctuates naturally between tool calls, sometimes dropping 30%+ between consecutive turns without any actual conversation truncation. This caused compaction counts of 30-40+ on sessions that never experienced a real compaction event.

Fix

Compaction now requires both a context drop and a hard cache-read drop, which indicates that the old cached prefix was invalidated:

if prev_context > 0
    && result.last_context_tokens < prev_context * 7 / 10
    && prev_cache_read > 1000
    && cr < prev_cache_read / 5
{
    compaction_count += 1;
}

The collector also preserves the previous context/cache state while tailing appended transcript bytes. On Unix, file identity now uses (device, inode) instead of mtime so normal appends stay on the incremental path; replacements still force a full reparse. Delta merges now carry forward prev_cache_read even when it drops to zero and add delta.compaction_count to the cached total.

Affected file

• src/collector/claude.rs — compaction detection, transcript file identity, incremental merge state, and regression coverage

Testing

• cargo test — 139 passed
• cargo clippy -- -D warnings — passed