[raymondginger2018-sudo]

Why

estimate_prompt_tokens and estimate_message_tokens decide when the pressure gate compacts or discards history, and history_signature is the contract by which the token meter's anchor and the compaction memo agree on what "unchanged" means. All three were exercised only indirectly, through the meter, so a silent change here would surface as a context overflow or an anchor that quietly stops anchoring.

What

New file tests/test_token_estimator.py (13 tests, no production code touched):

• Offline estimator arithmetic — locks len(payload) // 4 + 4 * messages and the 4-token floor, with the tokenizer forced off so every value is exact on any machine.
• Field collection — each field a tool turn can carry (tool_calls, reasoning_content, name, tool_call_id) and the tools schema is asserted to move the price; text-parts are priced like the equivalent string, non-text parts are not.
• CJK — the offline estimator is script-blind (equal length, equal price); under cl100k_base a CJK character costs more than ASCII. Both behaviours are pinned so a future script-aware change is a reviewed decision.
• Signature prefix property — the anchor keeps its sample only while new history starts with the signature it took; a shape/weight change that broke the prefix would silently drop every anchor.

Test plan

• pytest tests/test_token_estimator.py — 13 passed.
• pytest tests/test_token_meter.py tests/test_token_meter_wiring.py tests/test_tool_result_spill.py — all green (the tests that exercise these helpers indirectly).
• pre-commit run --all-files — all hooks passed.
• pip check — only unrelated ML-package conflicts from the local global env; CI installs from the lockfile in a clean venv.

[Zongwei9888]

Thanks @raymondginger2018-sudo, merged as 3d0bfe8. Having these helpers pinned directly is useful.

One note for later: the ("user", 0) in the prefix test reflects a gap in history_signature, which counts list content as zero length. If we fix that, we'll update the assertion with it.