[Xubqpanda]

Historical-source editors could stage an empty evidence directory, and DSH could launch with default model/provider/profile after credential filtering removed the resolved identity. This change keeps the selected source tree while supplying a validated run channel, and restores only the resolved non-secret DSH identity after filtering. Missing required identity or evidence now fails clearly; source-only calls can explicitly opt out of task evidence.

The shared editor APIs accept evidence_channel, with Dream-RSI passing its run channel without changing its selection rule. Bounded transcript logs now have truncation metadata and separate full-transcript usage receipts, per-call coverage, and parent/evidence/patch-base identifiers. Existing complete-transcript usage aggregation is retained; log truncation is not described as historical billing loss.

The real acceptance run exposed a ScienceWorld relay startup failure: importing its JAR constant eagerly imported an unavailable platform module. The digest import is now local to the platform-only payload lookup, with no new sandbox mounts or scoring changes.

Validation:

• 202 targeted offline tests passed, including real bubblewrap fake-CLI / fake-Node handoffs, non-default identity and credential isolation, historical/explicit/no-evidence routing, and isolated relay startup.
• 253 existing method-rule regressions passed (overlaps some targeted files).
• User-supplied offline reproducer confirmed both old defects against the unfixed checkout with zero model calls; the original package is preserved. Historical raw public archives were not in this package.
• The supplied optional Light evidence-navigation hook passed 9 offline tests and 4 real bubblewrap fake-CLI paths with the shared fixes. No Light gain was measured.
• Real acceptance uses original Loop 0.2.0, deployed Qwen3.5-9B, DSH + deepseek-v4.1-flash, 10 ScienceWorld training tasks, 4 workers, one improvement round. H0 completed 10/10, mean 0.147, with no invalid tasks. Dream-RSI exercised real DSH identity/evidence and usage receipts, but produced no source changes: H1 reused all 10 results. An unchanged SICA method additionally completed real edited-candidate scoring: 10/10 valid, mean 0.159, zero invalid/timeouts/harness failures, with ten native verdicts and ten traces. Two tasks improved, one regressed, seven stayed equal; platform content hashes matched before/after. Its real DSH session produced 26 usage receipts with matching resolved identity and parent/evidence/patch-base metadata. The candidate changes its turn/action budget (adaptive cap up to 96, up to six commands per reply), so its score is not a fixed-budget algorithm comparison.

The earlier failed acceptance attempt and the current run are retained as diagnostic evidence. Historical DREAM_RSI/Pencil and temporary-file issues are not assigned a guessed root cause. These are shared platform fixes, not Light algorithm gains. Task Harness sources and RSI selection rules are unchanged.