fix(benchmark): make concurrent checkpoint receiver admission reliable

[loopx-agent]

Concurrent receivers of one checkpoint could race while publishing the native bank member, leaving one arm failed before activation. Mature checkpoints could also pass launcher qualification but fail the installed receiver's default 10-second artifact check or 30-second native-history read.

This change serializes publication with a bounded per-point file lock, reuses the existing 60-second full-artifact allowance at workspace restoration, and shares a 90-second native admission allowance between initial context preparation and fresh-replan revalidation. Read allowances are capped by the original phase deadline; full history, cursor, identity, conflicting-bank and partial-attempt checks remain. Each receiver still obtains separate native run authority.

Stacked on #6622 (7b8b08f8a85dd7f0b87398d838af09f1b6a1bf03); this PR's base is its branch, not main. Apply the prerequisite stack in order or rebase and revalidate before merge.

Validation:

• Deterministic concurrent import regression fails on the parent and passes on the repair with the real native HTTP server and TypeScript artifact owner; grading is synthetic.
• Four targeted suites in Linux: 227 passed, 3 optional Docker smokes skipped; two source-packaging tests initially failed because the slim container lacked Git, then both passed after installing Git in a disposable container (229 passes total).
• Host adapter/handoff/shared-runtime suites: 146 passed, 43 dependency/live-runtime skips.
• A retained 65-turn checkpoint now passes full native admission/replay, 88-event Explore adoption and fresh-before-replan input preparation with the real native host. No model inference or grading was invoked by this qualification.
• A second retained point passes the production workspace restore (about 10 seconds), then fails at a separate Core adopt_todo_context full-artifact verification still using 10 seconds. That downstream gap is explicitly not closed here; do not claim this point's full receiver qualification or restart an outcome comparison from this evidence.
• Deadline expiry, short remaining allowance, symlink-lock rejection and bounded lock contention are covered. Public/private scan, compile and diff checks passed. Semantic advisory found no new shared vocabulary carriers; the full-tree semantic-vocabulary smoke passed.

Placement: existing benchmark native-provider transport and warm-receiver adapter; no checkpoint schema, scoring, opportunity, scheduler authority or cold-entry policy changes. The related refactor shares the native read/deadline logic across its two real callers. This is a bounded reliability repair, not sustained checkpoint-performance acceptance. CLI behavior is documented in the existing benchmark guide; no frontend or Lark surface changes.