chore(context): cut 7.6K chars of always-on rules

[paveg]

Summary

Second pass of the context-diet skill, this time on the always-on rules (the first pass, #244, covered skills and agent definitions). Effective always-on cost, after subtracting what the measure script cannot see (claudeMdExcludes, skillOverrides: off):

Class	Before	After
rules	26,215	18,613
skill descriptions	6,351	6,351
total chars per turn	32,566	24,964

Rules outweighed skill descriptions 4:1, so descriptions were left alone; the top three would save a few hundred chars for a full verification pass each.

Batches (one commit each, bisectable)

Batch	Before	After	What changed
zero-cost-scaling.md	1,749	0	Orphan. The rule became a skill in #212 but the deployed copy survived because deleting a source only makes chezmoi forget it. Added to .chezmoiremove; the skill body is byte-identical
workflow.md	7,154	3,617	Removed sections that restate development-principles.md, the paths-scoped tdd.md, agent frontmatter, or harness defaults. The 114K-token measurement and every worktree constraint stay verbatim
harness-engineering.md	3,078	1,962	The harness-vs-guardrail taxonomy is maintenance guidance, so it moved to harness-meta.md (paths-scoped to .claude/**). Every constraint survives, reworded
db-safety.md	1,560	1,319	Kept only the judgment the hook cannot make; dropped explanations of hook internals
markdown-formatting.md	2,913	2,027	Rationale paragraphs became one clause per bullet. The "When to diagram" table is unchanged because tests/writing-bench/rubric.md references it by name

Verification

• just test passes (lint, profile rendering, format, hook and skill-script tests).
• A skeptic subagent tried to refute the claim that every removed instruction is covered elsewhere. It found one factual regression (the hook description understated the quote/backslash normalization it performs) and two cross-references to a folded-away "Authoring rules" heading. Both are fixed in the last commit.
• chezmoi apply on this machine removed the orphan and deployed the new rules; chezmoi diff ~/.claude/rules is empty.
• Behavioral tests in fresh sessions were not run. The named-failure table for each cut lives in the local .ai/ note and is the regression list if something surfaces later.

Deliberately dropped, with no restatement anywhere

The skeptic flagged these five lines from workflow.md as lost. Each was cut on purpose because no concrete failure could be named for it, or because the harness already enforces it. Reviewers who want any of them back should say so here.

• "Use plan mode for verification steps, not just building"
• "Write detailed specs upfront to reduce ambiguity"
• The Demand Elegance re-prompt phrase ("Knowing everything I know now, implement the elegant solution"); the outcome is covered by the Simplicity ladder in development-principles.md
• "Track multi-step work with the harness Task tools; check in before starting implementation"; plan-mode approval already gates implementation
• The main session's own model tier ("opus; fable for the hardest"); this is session config, not something the agent acts on

🤖 Generated with Claude Code