[loopx-agent]

Benchmark checkpoint preflight still told an opted-in native solver to drain every page and reconstruct a delta against retained model input. The product's saved-file contract already allows reading the necessary current sections. Reuse the existing product navigation policy in the benchmark adapter, removing its separate eager-read and retention instructions. The original projection command and digest-bound full fallback are byte-for-byte unchanged; presentation remains delayed until the native tool runs after SessionStart.

The existing real-worker test now covers a progressive benchmark consumer as well as the original full-read consumer. Both exercise cold/warm wakes, original selection commands, the same native session, immutable source reconstruction and exactly one explicit inter-wake settlement. The adjacent App progressive path is retained. This uses native Codex with isolated SQLite and a deterministic loopback provider; it verifies transport and delivered instructions, not autonomous model judgment or accepted task outcomes.

Validation on 5609457ed86bf0cf73fd205a52efe512b23d0799 (base 3b4641c4fc0ec314d347f74b05880bb516012049):

• The existing preflight regression fails on the unchanged adapter and passes with the shared product policy. Capture/delivery suites: 101 passed, 7 native opt-in skips.
• Benchmark runtime/checkpoint suites: 147 passed, 64 existing opt-in/environment skips. The first standalone invocation lacked the repository test import root; the corrected invocation includes it and passes.
• Explicit native runs cover full-read benchmark, progressive benchmark and progressive App transport. Two initial progressive fixture probes used the wrong journal directory; corrected probes use the real isolated worker directory.
• Ruff, configured mypy (17 source files), diff/bilingual checks, semantic checks and the public-boundary scan pass. Standard premerge executes all nine selected canaries successfully, while retaining the benchmark_sensitive manual-review hold. No merge readiness is claimed.

For the fixed short-path builder fixture, initial prompt text grows from 1,285 to 1,714 UTF-8 bytes (+429) because it carries the complete existing navigation obligations. This removes an obsolete full-drain instruction, but does not establish net input, wall-time, fee or long-horizon benefit. Default-off behavior, scoring, case semantics, budgets, admission and settlement rules are unchanged. No paid inference, managed benchmark campaign, installation or model configuration change is included.

The bounded refactor replaces a caller-owned policy with the existing context owner. It adds no protocol field, cache, reader, lifecycle owner or UI. C1 cutover (#7430), numeric fidelity (#7661), actual solver adoption and complete outcome/cost qualification remain with their existing owners.