Render enriched subagent progress in the run-once console

[rochecompaan]

Parent: #116

Depends on #125.

Outcome

Run-once operators see at least one concise progress outcome per subagent child: one line for each distinct authoritative agent/model/thinking tuple, or one safe agent-only fallback when no authoritative tuple arrives.

Desired output

🤖 subagent (agent=reviewer, model=gpt-5.6-sol, thinking=xhigh)

Scope

• Render available agent, model, and thinking fields as comma-separated values in the existing 🤖 subagent (...) format.
• Preserve model identifiers containing nested provider or model path segments.
• Omit unavailable fields rather than inventing values.
• Preserve existing management-call and non-child subagent formatting.
• Keep progress on stderr and final run-once JSON on stdout.
• Add focused console and end-to-end regression coverage.

Out of scope

• Metadata resolution or inference.
• Pi lifecycle observation.
• Session streaming, cancellation, child inventory, or accounting.
• Changing pi-subagents TUI output.
• Enriching triage output.

Acceptance criteria

• Each distinct authoritative tuple renders one stable line with every available field, while exact duplicates are suppressed.
• A changed authoritative tuple renders an additional line without replacing earlier output.
• Missing model or thinking fields are omitted cleanly.
• Nested model identifiers are preserved byte-for-byte.
• Failed and unresolved children remain visible through the fallback contract from Correlate per-child subagent progress and preserve pipeline accounting #125.
• Stdout remains machine-readable final JSON and progress remains on stderr.
• Existing run-once output outside child execution is unchanged.

Planning constraints

This issue owns one focused console/output PR. Its spec and plan supersede only the rendering and final end-to-end verification portions of the deprecated monolithic plan referenced by #116.