fix(compression): count Codex replay fields in tail-protection budget (#55572)

[teknium1]

Summary

Codex/Responses sessions no longer compact continuously near the 272K cap: the tail-protection token estimator now counts provider replay fields (codex_reasoning_items, codex_message_items, reasoning, reasoning_content, reasoning_details), so compaction actually sheds enough tokens to get back under threshold.

Root cause: the preflight "should I compress?" estimator stringifies the full message dict and sees the encrypted reasoning blobs, but _estimate_msg_budget_tokens (the tail-protection walk) only counted content + tool_calls. On a real 214-turn Codex session those blobs were ~115K tokens (27% of the payload) — the "20K" protected tail silently held 100K+ real wire tokens, so each compaction left the session above threshold and re-fired. Fixes #55572; resolves the community report of gpt-5.5 sessions becoming unusable at ~260K.

Salvaged from PR #45485 by @bigstar0920 (cherry-picked, authorship preserved), rebased onto current main: the helper is folded into the existing _estimate_msg_budget_tokens (which #28053's fix had already centralized since the PR branched) so all four call sites — both _find_tail_cut_by_tokens walks and the tool-result prune boundary walk — get the replay accounting.

Changes

• agent/context_compressor.py: _serialized_length_for_budget() helper + _REPLAY_BUDGET_KEYS; _estimate_msg_budget_tokens now adds replay-field weight (accounting only, never mutates/prunes the fields)
• tests/agent/test_context_compressor.py: contributor's TestTailBudgetCodexReplayFields suite (per-field parametrized + combined case)

Validation

	Before	After
Tail estimate, 10-msg Codex tail w/ 8KB blobs	765 tok	10,825 tok (real weight)
E2E compress() shed on 60-turn Codex-like session	near-zero (tail held blobs)	67.7% serialized chars
Blob-bearing messages kept	all in tail	19/60 (tail only)
tests: test_context_compressor.py, compression_boundary, infinite_compaction_loop, image_tokens	—	all pass

E2E run via real ContextCompressor.compress() with mocked LLM against temp HERMES_HOME.

Infographic