fix: deliver inter-step output via beads scope, not inlined into the dispatch (millworks-c30)

[richardkiene]

The bug (found in real project use)

A workflow step's task template interpolates upstream outputs via {step.X.output}/{previous_output}. substituteVariables inlined the entire upstream output (a ~10KB requirements doc) into the task string, and the Claude dispatcher then typed that whole string into the pane via tmux send-keys -l — which blew past send-keys' length ceiling, so the dispatch failed before the subagent started. It's a ceiling on inter-step payload size: every downstream step (architecture, optimization, code-gen) embeds the same doc and fails identically. pi dodged it only via a wrapper-file argv (higher ARG_MAX ceiling, same inline smell).

The output is already in beads (STEP notes, since inc5) — so re-serializing it through a keystroke channel was both redundant and unscalable.

The fix (lockstep, beads-native)

Deliver the upstream output through the beads-scoped context bundle (a file the assembler builds) instead of inlining it into the typed/argv task:

• {step.X.output}/{previous_output} resolve to a short reference (stepOutputRef) — identical on both surfaces — keeping the author's intent without the payload. Validation (dependsOn membership, settled) is unchanged.
• The dependency steps' beads records are scoped in at dispatch (pi dispatchStep; Claude threads beadsScopeIds through assembleContext → assembleContextViaCli → --beads-scope, which the Claude surface never passed before).
• The assembler surfaces the deps' output: run_bd_show was parsing bd show --json as an object (it returns an array) and reading a nonexistent body field capped at 3 lines, so beads-scoped records rendered ~nothing. Split into a pure, array-aware summarize_bd_record that surfaces the full STEP notes under a step:<id> heading. Large notes are budget-pruned by the assembler's existing 80% pruner — a graceful cap, vs the send-keys hard fail.

Net: beads is the source the data flows from; the typed/argv task shrinks to the instruction; no payload through send-keys/ARG_MAX.

Testing

• 4 new Rust unit tests for summarize_bd_record (array, full notes, step-label, fallback, none-cases).
• Both surfaces' substitution + drive tests updated to the reference contract; pi 128 + Claude 270 green, typecheck clean.
• Verified end-to-end against real bd: the built context-pack-assembler --beads-scope <stepid> surfaces the step's notes labeled by step:<id> in the bundle.
• Remaining: live verification in the blocked greenfield-compile run (resume past requirements→feasibility) after a plugin rebuild — owner to run.

Notes

• Lockstep across pi + Claude + the shared Rust assembler (a run from either surface stays read-back-compatible).
• The deeper "steps emit structured DECISION/RISK/REQ records as canonical output" direction is scoped separately as epic millworks-cn8.
• Pre-existing context-pack-assembler bd-prime test fragility (millworks-rrp) is untouched and unrelated.