[Parent] Track self-optimization design follow-ups (extracted from #1061) #2076

Umbrella tracking issue for the five follow-up items extracted from the closed design PR #1061 ("feat(compile): IR-validated step authoring & runtime self-optimization"). Each of these captures a piece of that design that was deliberately deferred from the initial landing and needs its own focused implementation:

• Hoist deterministic repeated agent work into workflow steps #2067 — Design a current-architecture mechanism for hoisting deterministic, repeatedly observed agent work into generated steps/post-steps, building on the typed IR and generated-shell binding model.
• Add strict validation for untrusted raw step fragments #2068 — Add a strict validation path for isolated untrusted raw-YAML step fragments (Step::RawYaml), using the current typed task parsers and lint findings.
• Add preview-only self-optimization safe output #2069 — Introduce an opt-in, preview-only self-optimization proposal flow using the current configured-tool and safe-output architecture (no source mutation in v1).
• Ground optimization proposals in recorded telemetry #2070 — Add programmatic verification that a proposed optimization is supported by recorded execution telemetry, rather than relying on agent-supplied evidence alone.
• Preserve comments when updating workflow front matter #2071 — Implement or adopt a comment- and formatting-preserving front-matter editor, required before any automated workflow-source update is supported.

Together these define the safe, staged path toward the "runtime self-optimization" capability: hoisting deterministic work, validating untrusted fragments, telemetry-grounded proposals, and finally an editor safe enough to auto-apply them.

Closing this parent when all five sub-issues are resolved (or explicitly deferred) gives one point of visibility into this follow-up set.

Generated by Issue Arborist.

Generated by Issue Arborist · auto · 78.6 AIC · ⌖ 2.02 AIC · ⊞ 9.7K · ◷