Lower prompt regression budgets after the v1.7 rewrites

[jrpbuilds]

Parent

Part of #46.

Relationships

Depends on #57, #58, #59, #60, #61 and #62.

Goal

Lock in the v1.7 prompt simplification with aggressive regression budgets so future features cannot quietly rebuild the large procedural prompts this project removes.

The budget is a bloat tripwire, not a target to fill.

Scope

Update the existing prompt-budget tests after the role rewrites have stabilized. Measure expanded/runtime prompt content in the same form agents actually receive where practical, not only raw source-file size if includes materially change the result.

Use final measured v1.7 sizes to set limits with modest headroom.

Initial design targets to guide the rewrite are approximately:

Prompt	Initial target
Coordinator base	< 10 KB
Coordinator mode fragment	< 3 KB
Explorer	< 3 KB
Planner	< 6 KB
Designer	< 5 KB
Implementer	< 6 KB
Reviewer	< 10 KB
Review critics	< 3 KB each
Frontier	< 3 KB

These are deliberately rough. If a role can be substantially smaller without losing behaviour, set the budget closer to the actual result rather than the table ceiling.

Budget policy

• Reviewer may legitimately remain larger than simpler specialists because its output protocol is semantically richer.
• A budget increase in future should require an intentional test/code change and explanation.
• Do not game budgets by moving equivalent always-included prose into another file.
• Consider a total expanded-prompt budget as an additional regression check if it catches cross-fragment growth better than per-file checks alone.

Acceptance criteria

• All major role prompts have materially lower v1.7 regression limits.
• Limits are based on final rewritten prompt sizes with small justified headroom.
• Include-expanded content is accounted for where relevant.
• A total prompt-size regression check is added if useful.
• Budget tests fail clearly with role/current size/allowed size.
• CI makes future prompt growth an explicit engineering decision.
• No semantic behaviour is removed solely to satisfy an arbitrary number.

Out of scope

• Tokenizer-specific billing estimates unless trivial to add.
• Benchmarking run latency/cost; tracked separately.