Feature Request: $LOOP_PREV_OUTPUT variable for loop nodes

Problem

• loop: nodes in DAG workflows have no way to inject the previous iteration's output into the next iteration's prompt when fresh_context: true. The variable lastIterationOutput is computed in executeLoopNode at the end of each iteration but is immediately discarded — it is never exposed to the prompt substitution system.
• Experienced by anyone authoring loop: nodes that use fresh_context: true and need the AI to know why the previous pass failed (e.g. QA failure summaries, validation errors, review feedback) without carrying the full conversation history forward.
• Comes up in every retry-on-failure pattern: implement→validate loops, review→fix loops, generation→evaluation loops.

Proposed Solution

Add a $LOOP_PREV_OUTPUT substitution variable populated from the previous iteration's cleaned output, available inside loop.prompt from iteration 2 onward. On iteration 1 it substitutes to an empty string. When fresh_context: false the variable still works but is redundant since session history already carries the context.

Implementation touches two places:

• executor-shared.ts — add loopPrevOutput?: string parameter to substituteWorkflowVariables
• dag-executor.ts (executeLoopNode) — pass lastIterationOutput (already computed at line 1751) into the substitution call on each subsequent iteration

User Flow

Before (current)

loop:
  prompt: |
    Implement the plan, then run `bun run validate`.
    If checks fail, fix them and retry.
    [!] No way to reference why the previous iteration failed —
    [!] the agent starts each fresh-context iteration blind.
    When all checks pass output: <promise>QA_PASS</promise>
  until: QA_PASS
  fresh_context: true
  max_iterations: 3

After (proposed)

loop:
  prompt: |
    Implement the plan, then run `bun run validate`.

    [+] Previous iteration output (empty on first pass):
    [+] $LOOP_PREV_OUTPUT

    Use the above to understand what failed last time (if anything)
    and focus your fixes there. When all checks pass output: <promise>QA_PASS</promise>
  until: QA_PASS
  fresh_context: true
  max_iterations: 3

Alternatives Considered

Scope

• Package(s) likely affected: workflows
• Breaking change? No — $LOOP_PREV_OUTPUT substitutes to empty string when absent/first iteration; existing prompts that don't reference it are unaffected
• Database changes needed? No
• New external dependencies? No

Security Considerations

• New permissions/capabilities? No
• New external network calls? No
• Secrets/tokens handling? No — $LOOP_PREV_OUTPUT reflects AI-generated output from within the same workflow run, same trust boundary as all other $nodeId.output substitutions

Definition of Done

• substituteWorkflowVariables in executor-shared.ts accepts and substitutes $LOOP_PREV_OUTPUT
• executeLoopNode passes lastIterationOutput (empty string on iteration 1) to substituteWorkflowVariables on every iteration
• $LOOP_PREV_OUTPUT documented alongside $LOOP_USER_INPUT in CLAUDE.md variable reference
• archon-compose-plan-implement-qa.yaml updated to demonstrate the variable in its prompt
• Unit test: $LOOP_PREV_OUTPUT is empty on iteration 1, equals previous output on iteration 2+
• Unit test: prompts that do not reference $LOOP_PREV_OUTPUT are unaffected