fix(agent): auto-compact no longer silently drops head content past the 80K slice

[shadowinlife]

Summary

• Fixes the silent information loss in AgentLoop._auto_compact() reported in [Bug] Auto-compaction can silently drop conversation content during summarization #1055: the head is now folded through the summarization LLM in budget-sized, message-boundary chunks instead of being hard-sliced at 80,000 serialized characters.
• Every head message now reaches the summarization LLM exactly once, upholding the method's documented "zero info decay" contract.

Why

_auto_compact() built its summarization prompt with:

conv_text = json.dumps(head, default=str, ensure_ascii=False)[:80000]

head is everything not kept in the ~20K-token tail (TAIL_TOKEN_BUDGET). Once json.dumps(head) exceeded 80,000 characters, everything past the cut was neither seen by the summarization LLM nor preserved in the tail — silent information loss with no exception, contradicting the docstring's "zero info decay" guarantee.

Verified against main @ c33133f with the issue's reproduction script: messages [7..13] were present in neither the summarization prompt nor the reconstructed conversation.

Closes #1055

Changes

• agent/src/agent/loop.py
  • New module helper _chunk_messages_for_summary(): splits messages into chunks whose serialized JSON stays under the budget; message boundaries are never broken; a single oversized message forms its own chunk rather than being dropped.
  • New constant COMPACT_HEAD_CHAR_BUDGET = 80_000 (same per-pass budget as the old slice).
  • _auto_compact() replaces the single truncated pass with a fold: the first pass starts from the previous summary (or the structured template, keeping prior behavior), and each subsequent pass folds the next chunk into the running summary via the existing _ITERATIVE_UPDATE_PROMPT. Single-chunk heads behave exactly as before.
  • Docstring updated to state the chunked no-drop guarantee.
• agent/tests/test_auto_compact_no_silent_loss.py — regression tests:
  • the issue's exact scenario (20 tagged messages, oversized head): zero lost markers, multiple summarization passes;
  • chunk helper: exact-once coverage, per-chunk budget invariant, oversized-single-message preservation, empty input.

Note for maintainers: this touches the protected src/agent/ path. The issue author asked which remediation direction was intended (overflow preservation vs. chunked summarization). This PR implements chunked summarization, since it is the only option that keeps the "every message is either summarized or preserved verbatim" contract without growing the retained tail unboundedly; happy to reshape if a different direction is preferred.

Test Plan

• Existing tests pass (pytest --ignore=agent/tests/e2e_backtest --tb=short -q) — 10079 passed, 76 skipped
• New tests added — 4 regression tests in agent/tests/test_auto_compact_no_silent_loss.py
• Tested manually — issue [Bug] Auto-compaction can silently drop conversation content during summarization #1055's reproduction script: before the fix, 7-8 markers lost; after the fix, 3 summarization passes and zero lost markers

Checklist

• No changes to protected areas (src/agent/, src/session/, src/providers/) without prior discussion — touches src/agent/loop.py; flagged above for maintainer direction, per the question raised in [Bug] Auto-compaction can silently drop conversation content during summarization #1055
• No hardcoded values (API keys, file paths, magic numbers) — the 80K budget is a named constant
• Code follows CONTRIBUTING.md guidelines — DCO sign-off on the commit
• Documentation updated (if user-facing change) — _auto_compact docstring updated