feat(launcher): re-stage source_code targets when Soundwave replans

[Lempkey]

Problem Statement

The stageSourceTargets goroutine in the launcher is one-shot: it polls for plan/.bundle_complete, stages all source_code targets once, then exits. If the operator asks Soundwave to revise the engagement plan — changing the target path, adding a new source target, or removing one — complete_engagement_planning will write a fresh .bundle_complete marker with updated contents. The goroutine is already gone and will never observe the new marker.

The result is that /workspace/src inside the sandbox reflects the original plan, not the replanned one. The analyst continues with the wrong source tree, and there is no warning to the operator. This is a silent correctness failure for any engagement that goes through a planning revision.

Proposed Solution

Replace the one-shot goroutine with a persistent watcher that survives the initial staging run and continues to observe the plan directory for marker refreshes.

Concretely:

1. After a successful staging run, instead of returning, the goroutine re-enters its polling loop watching for a new .bundle_complete whose completed_at timestamp is newer than the one it last processed
2. When a newer marker is detected (and the mtime consistency check passes), re-read roe.json, diff the source targets against the previously staged set, and re-run copyDirTree for any targets that changed
3. Write staging-status.json back to in_progress during the re-stage, then back to staged on completion
4. Print an operator-visible note in the CLI: "Replan detected — re-staging source targets..."

The goroutine should exit only when the engagement workspace is torn down (context cancellation / launcher shutdown).

Alternative (simpler): Re-launch the goroutine from runStart each time a new engagement_ready event is observed on the stream. This avoids persistent state in the goroutine at the cost of always re-copying the full tree (no diffing).

Alternatives Considered

• Do nothing / document the limitation: Operators who replan must restart the launcher to pick up new source targets. Acceptable short-term but a footgun for iterative engagements.
• Full re-copy on every replan: Simpler implementation than diffing; acceptable for most source trees since staging is a local copy and typically fast relative to planning time.
• Operator warning only: Detect a stale /workspace/src (by comparing manifest against current roe.json) at objective dispatch time and warn the analyst. Does not fix the root cause but prevents silent wrong-tree analysis.

Area

Docker / Infrastructure

Additional Context

Identified as a pre-existing architectural gap during the Codex adversarial review of PR #203. The one-shot design was intentional in the original implementation to keep the goroutine simple, but the introduction of plan/.bundle_complete in PR #203 makes it straightforward to detect replans by watching for a marker whose completed_at is newer than the last-processed timestamp.

The staging-status.json manifest written by PR #203 already records which source paths were staged and when, providing the baseline needed for a diff-based re-stage.

Related: #202, PR #203, #204

[PurpleCHOIms]

Closing as moot — see PR #203 and #202 close comments. The one-shot staging design this issue addressed is being declined alongside the rest of the staging chain.