Add complexity/label-based model or effort tiering for the code agent in .fullsend/config.yaml

What happened

Issue #131 (https://github.com/redhat-et/ProtoBot/issues/131) requested a purely mechanical, 7-line documentation fix (adding a Contents/TOC block to one ADR file, mirroring an existing sibling format). Triage labeled it documentation and confirmed it was low-complexity. The code agent run (https://github.com/redhat-et/ProtoBot/actions/runs/35376105082) nonetheless ran at high effort on xai/grok-4.6 per .fullsend/config.yaml, costing $2.66 and taking ~12 minutes — about 18x the triage run's cost ($0.15) and roughly 2x the review run's cost ($1.26), for a change that ultimately touched a single file with no logic, interfaces, or cross-document consistency at stake. Inspection of .fullsend/config.yaml confirms the code agent's model is hardcoded (xai/grok-4.6) with no branch based on issue labels, estimated diff size, or triage-assigned complexity.

What could go better

This is one data point, but the root cause is structural and directly observable in the config, not incidental: there is currently no mechanism in .fullsend/config.yaml to scale the code agent's model or effort down for issues that triage has already identified as low-complexity (e.g., via the documentation label or an explicit complexity signal). This means every documentation-only or trivially-scoped issue pays the same premium-model cost as a substantive code change. I'm confident the cost skew is real (it's a direct dollar comparison from the run metadata); I'm less certain what the ideal tiering rule should be (label-based vs. triage-emitted complexity score vs. diff-size heuristic post-hoc) — that design choice is left to the implementer.

Proposed change

In redhat-et/ProtoBot's .fullsend/config.yaml, add a routing rule so the code stage selects a lower-cost model/effort tier (e.g., google-vertex/gemini-3.8-flash, already used elsewhere in this config for fix/prioritize) when the triggering issue carries the documentation label (or a future triage-emitted complexity signal), falling back to the current high-effort grok-4.6 tier for everything else. Keep the review stage's existing model mix unchanged since it already varies by sub-agent role.

Validation criteria

Over the next 10 code-agent runs on issues labeled documentation, average per-run cost should drop measurably (e.g., below ~$0.75, roughly in line with the review-agent cost for this trivial change) without any increase in review-agent-requested-changes or human-requested-changes rate on those same PRs compared to the current baseline.

Generated by retro agent from https://github.com/redhat-et/ProtoBot/pull/146

Start by reading .fullsend/config.yaml, especially the code stage and the existing model choices used by fix/prioritize; review the referenced issue #131 and run metadata for the baseline. Done means documentation-labeled runs use a lower-cost code-agent tier, other runs retain the current tier, review settings remain unchanged, and the next 10 documentation runs show lower average cost without higher change-request rates.