emrg: add unit tests for daemon token estimation and history serialization helpers

[argszero]

Summary

Adds 22 new unit tests for daemon's pure-function helpers that were previously untested:

Function	Tests	What's covered
_count_chars_for_tokens	5	Pure ASCII, pure CJK, mixed, empty, Kana
_estimate_tokens	4	Empty list, single message, tool_calls, multiple messages
_estimate_single	3	With content, empty content, non-string content
_records_to_text	6	Empty, message, tool_call, tool_result (500-char truncation), summary, mixed
_truncate_record	4	Short pass-through, long truncation, exact boundary, field preservation

All tests are pure-function, no mocking needed. From 15 tests → 37 tests in test_daemon.py.

Why

These functions are critical for the auto-compact flow — token overestimation causes premature compaction, underestimation causes context overflow. Having tests ensures future refactoring (tasks.yml migration) won't silently break token estimation.

Verification

319 passed in 0.63s

[argszero]

✅ LGTM — cycle #4

[argszero]

✅ LGTM — cycle #4 (37 tests pass, good daemon + scheduler coverage)

[argszero]

✅ LGTM — cycle #5 (daemon tests solid, token estimation accurate)

[argszero]

✅ LGTM — cycle #6 (389 pass, clean diff after rebase)