feat(kanban): harden budget-aware execution recovery

[GravityTone]

Summary

This PR hardens Kanban execution so long-running work can use fewer context tokens and iterations without increasing max_turns.

• adds durable iteration-budget checkpoints and exactly-once compression/repair behavior
• makes compact Kanban context bounded to 48 KiB UTF-8 while reserving LATEST_HANDOFF and CONTEXT_REF
• keeps handoffs reversible through hash-bound recovery of partial_summary_full
• adds run CAS/ownership checks so stale workers cannot mutate a newer run
• makes exact-review provenance durable and fail-closed across direct creation and decomposition
• constrains decomposition to atomic 2–6 child fanout and blocks hidden activation/release work as PREPARE_ONLY
• strips private Hermes metadata before provider sizing and wire serialization, while preserving repaired user input exactly once
• documents configuration and ships a reproducible synthetic context benchmark

Builds on the timeout-salvage work in #52511 and implements the preventive budget-warning contract from #54153. The original contribution remains credited in the commit trailer.

Verification

Verification on the final integrated candidate:

• Kanban matrix plus benchmark tests: 822 passed
• agent/run-loop, 413 recovery, checkpoint, role-repair, compression, streaming, multimodal, and Bedrock tests: 782 passed
• configuration matrix: 300 passed
• combined canonical run: 1,904 passed, 0 failed, subprocess return code 0
• Ruff, git diff --check, Python 3.11 compile, and focused Python 3.9 compile: PASS
• YAML/JSON parsing, deterministic benchmark, and security/privacy scans: PASS

Synthetic benchmark:

• full context: 340,866 bytes
• compact context: 49,152 bytes
• reduction: 85.58%
• repeated and versioned artifacts: byte-identical

Review receipts

• final candidate: efa7a991148be122e93f7c898f5adf98d4f55083
• base reviewed for overlapping integration: c78aa0bad5dc96c8080b1d16def868f1ab039b0c
• tree: b6b04c8eaa3be35c4a6d895d7d2e6f6b5dfefa33
• binary diff SHA-256: 9f3dea8ac55998cf9302ccab61a55c3c5a47e4b9843de4e86095121ee2122d90
• bundle SHA-256: 1cbf19e12bc68b1a515706ce9ae8f0883f5c05409c292666ae8447ede02fd55e
• independent material controller: PASS
• overlapping-integration controller: PASS
• canonical PRE/POST interlock: PASS / NO_MUTATION
• candidate history: one privacy-safe squash commit

After the frozen review, main advanced by four commits touching three paths. The intersection with the feature's 36 paths is empty; the reviewed base remains an ancestor, and git merge-tree reports a clean merge against current main.

[teknium1]

Thanks for the substantial recovery and ownership hardening work. The current main path still lacks the run-id guard in agent/turn_finalizer.py:109-131, so the underlying Kanban recovery premise remains valid.

Problems

• Blocking: agent/agent_runtime_helpers.py:422-425 now removes every underscore-prefixed message field before the provider path. The PR uses that copy for every request in agent/conversation_loop.py:979. This removes _anthropic_content_blocks, but current agent/anthropic_adapter.py:2069-2077 explicitly reads that field to preserve legacy stashed multimodal tool-result blocks. Anthropic requests are built from these transformed messages via agent/chat_completion_helpers.py:859-869, so the compatibility path loses its image/content blocks.

Suggested changes

• Preserve _anthropic_content_blocks until Anthropic conversion, or restrict early stripping to known bookkeeping markers and retain generic transport sanitization afterward.
• Add an Anthropic request-path regression test for a tool message containing _anthropic_content_blocks.

Automated hermes-sweeper review.

[teknium1]

Blocking: this blanket underscore-key removal also drops _anthropic_content_blocks. The Anthropic converter explicitly consumes that back-compat field in agent/anthropic_adapter.py:2069-2077, and this helper is now used before provider conversion at agent/conversation_loop.py:979. Preserve that field until Anthropic conversion (or narrow this filter) and add a request-path regression test.