benchmark: aggregate EdgeBench fix round for the next 24h run

[songoow]

Purpose

Prepare one branch for the next EdgeBench experiment. It combines refused-checkpoint handoff resilience, bounded full capture preparation, initial-delivery guidance and evidence-based research followthrough.

A refused handoff keeps the original wake running and records a bounded diagnostic. Full capture preparation shares a 90-second window bounded by the original trial phase. Planning compares a feasible candidate with the strongest available starting point and separates informative evaluation from replacing the best artifact.

Research followthrough

Stride reviews return the prior decision, its declared Todo targets, canonical work state and later observations. Outcome citations are replayed against complete scoped history at writeback: display caps and an omitted optional typed-progress attachment cannot invalidate real work evidence. Old or unrelated evidence cannot establish execution of the declared work. Negative and unknown outcomes remain legal.

Legacy descriptive Todo deltas remain readable with unresolved linkage; overlong references cannot alias another Todo. Structured start packets reuse the shared planning guidance, while standalone messages retain the full instruction. No new benchmark acceptance switch or score threshold is introduced; stride remains opt-in and historical decision identities are preserved.

Validation

Focused validation for the new changes:

• 40 TypeScript replan/stride tests passed, including capped history, ordinary work evidence and legacy declarations.
• 2 isolated real File/SQLite CLI continuation tests passed, including negative-result review and rejected old-evidence recovery.
• 2 planning-packet/standalone-message checks and the focused output-guidance recognition check passed.
• Control-plane TypeScript typecheck, focused Python lint and diff checks passed.
• Structured guidance copies reduced from four/six to one/two in compact/detailed guided packets. Standalone messages remain self-contained; the bounded first-introduction allowance documents their retained content.

Full CI was not awaited. No live model inference, benchmark launch or score-improvement claim is part of this change. Real model adoption, sustained operation and research benefit remain to be measured. Packaged frontend adoption of the added review detail is not qualified here.

This remains an experiment-preparation branch; it has not been merged into main. The original checkpoint and first-delivery components remain traceable to #6618 and #6621.

Branch repairs after the first launch attempt (fb6f060, a0fa279)

A first paired launch on head f7756ff lost every LoopX arm within two minutes on both hosts: RuntimeError("Native task prompt is unavailable") in SForgeWorker.format_run_cmd, before the first turn. The provider (decf67c) and the workspace-instruction install were unchanged; the discriminator against the earlier run that completed 11 rounds was the launcher's umask 077. The worker re-uploads the feedback prompt from a host temp file, the native copy preserves that mode and lands the file root-owned, so a 0600 prompt was unreadable by the container's agent user. a0fa27981 sets the prompt mode explicitly before upload and adds a regression test that runs format_run_cmd under umask 077 (fails on the previous head with mode 0600, passes now). The checkpoint declaration upload was already chowned and re-moded after upload and is unaffected.

fb6f060e7 repairs the branch's own kernel-static-checks failure. The base/head differential copies cli_output_semantics.py out of the package and measures a base checkout that predates the planner guidance constant: the relative import had no parent, and an absolute import fails on that base. A missing owner is now observed as "guidance absent". The bootstrap JSON pack carries its rendered message, so the expanded guidance lands twice there (planner field plus the labelled block message-only consumers need); measured one-time growth is 2,821 chars, and that surface's allowance is raised to 3,000 with the measured value recorded in the parity fixture. Other surfaces keep their limits.

Validation on a0fa279 (source checkout, uv run --extra test): cli-output-budget-regression-smoke ok; 176 cli_output control-plane tests; 49 EdgeBench prompt tests including the new umask case; 322 SForge runtime, checkpoint-handoff and warm-start tests; 40 TypeScript replan/stride tests. Full CI on this head was not awaited at the time of writing.

Planning-wake retry (b7c47c9)

The second paired launch on f7756ff got past the prompt but every LoopX arm still ended after one wake: the model stream disconnected during the planning wake (eight failed reconnects, turn.failed), the entry raised "Planning failed; execution was not started", and the trial closed with no submission. Explore activation, stride and the checkpoint declaration had all succeeded. prepare_entry now retries a planning wake whose native process exited non-zero without a plan or a harness error, up to three attempts inside the same phase budget, and records every attempt in planning-attempts.json beside the planning receipt. Timeouts, harness-side failures, a zero exit without a plan and an exhausted budget still fail once; a validated or blocked plan is never re-run. This is a default harness behavior change for the loopx-planned entry and is covered by three new entry tests (58 task-entry tests, 323 runtime/handoff/prompt tests pass locally).