feat(rules): add intent-discipline rule package

[Mathews-Tom]

Summary

Adds a new rule-type package, intent-discipline, that codifies four behavioral guardrails for non-trivial coding work. It targets the failure modes where an agent runs ahead of the user's actual intent: silently picking one interpretation of an ambiguous request, over-engineering single-use code, mixing drive-by refactors into an unrelated diff, and reporting work "done" against a success criterion too weak to verify.

The rule is deliberately distilled against the existing armory rule set so it adds discipline that no current rule covers, and cross-references the rules it borders rather than restating them.

The four guardrails

1. Surface intent before implementation — state assumptions, present multiple interpretations for ambiguous requests, push back when a simpler approach exists, stop and ask when confused.
2. Minimum viable implementation — smallest code that solves the stated problem; no speculative abstractions for single-use code; the "200 → 50 line" simplicity test.
3. Surgical changes — every added/removed/modified line traces to the request; no adjacent "improvements" or style drift; remove only the orphans your own change created.
4. Verifiable goals over imperatives — convert "fix it" / "make it faster" into executable success criteria the agent can loop against; multi-step plans carry a non-optional verify column.

Distillation against existing rules

The rule explicitly defers instead of duplicating:

Concern	Owned by
Test-first execution mechanics, coverage thresholds	test-standards (cross-referenced)
Commit message format, diff-per-commit hygiene	commit-standards (cross-referenced)
Token-efficient tool calls	token-efficiency (out of scope)
Reasoning-depth control	adaptive-thinking-control (out of scope)

A When to Use This Rule vs Related Rules table in the rule body draws these boundaries for the reader.

Hardening

Beyond the source framework, the rule adds:

• When NOT to Apply — explicit trivial-task escape hatch (typo fixes, one-line tweaks, throwaway scripts) so the caution bias does not tax simple work.
• Rationalizations to Reject — six common agent excuses paired with factual rebuttals.
• Red Flags — observable patterns indicating the rule is being violated (unrequested file changes, code-block-first responses to ambiguous prompts, single-implementation abstractions, "should work" verification).

Changes

• rules/intent-discipline/RULE.md — rule definition (frontmatter description trimmed to fit the 1024-char validator ceiling; triggers span the intent / simplicity / verifiability synonym families).
• rules/intent-discipline/evals/cases.yaml — five positive cases (ambiguous request, over-engineering, drive-by refactor, imperative-to-verifiable, plan-with-verification) and three negative cases that disambiguate from typing-standards, test-standards, and commit-standards.
• manifest.yaml — regenerated; rule count 5 → 6, total packages 130 → 131.
• README.md — rule added to the Rules table; package count badge 130 → 131.
• ATTRIBUTIONS.md — framework origin recorded under Conceptual Inspiration.

Commits

Two logical, bisectable commits:

• feat(rules): add intent-discipline rule package — rule + evals + manifest (manifest travels with the package so the rule is self-consistent and CI-valid at this commit alone).
• docs: surface intent-discipline in catalog and attributions — README + ATTRIBUTIONS.

Validation

Check	Result
validate_frontmatter.py	131/131 packages
validate_evals.py	all 6 rules
validate_references.py	131 packages, references intact
sync_templates.py	in sync
Bisect check	commit 1 alone passes all three validators

Cursor / Codex / Gemini adapters regenerate from the rule via generate_adapters.py (gitignored build artifacts, not committed).

Test plan

• CI: manifest sync + eval validation pass
• Run package-evaluator against rules/intent-discipline/ — target ≥70% (Adequate), aim ≥80% (Strong)
• Confirm the Rules table renders and the package count badge reads 131