Restore structural rules dropped in 780a7e8 (efficiency floor, danger-seam map, +4)

[oddyblue]

What

Adds six verified-lost structural rules on top of 780a7e8 ("Make subtraction and scope closure enforceable"), leaving that commit's enforceability gains fully intact:

• Efficiency floor — "no gratuitous recomputation, repeated I/O, or quadratic passes over unbounded input, and never contort readable code for an unmeasured micro-win." Nothing currently guards a "simplification" that introduces an O(n²) pass.
• Danger-seam map — "look hardest where simplification tends to break things": state synchronization; concurrency, async lifetimes, and cancellation; navigation, lifecycle, and scheduling; storage schemas, migrations, and caching; configuration and environment branches; permissions and external integrations. (780a7e8 had dropped caching, permissions, navigation, and scheduling entirely.)
• Naming-as-probe — "a thing you cannot name honestly is usually the wrong shape."
• Actual-date anchor — verify drift-prone facts "as of today's actual date" (counters an agent reasoning from its training cutoff).
• Competing-explanations — "when diagnosing, weigh the competing explanations for the cause, not only the candidate fixes."
• Structural cleanups — the 7-step workflow re-homed under an accurate heading (Work the owning path, not the diff window) instead of sitting under "Establish the real problem"; the three mirror bodies re-synced (header whitespace).

All three mirror bodies (SKILL.md, AGENTS.md, .cursor/rules/less-is-more.mdc) remain byte-identical.

Why

A close review of 780a7e8 confirmed it made subtraction and scope-closure genuinely enforceable — those gains are kept here in full. But it also dropped several items that serve the skill's stated purpose. This PR restores the low-risk structural ones (additive rules, not exhortations).

Deliberately excluded (a separate, doctrinal decision)

The counterweight ("a complete fix, not a timid half-measure"; "the bar cuts both ways") and the purpose-framing ("the best solution the problem justifies"; "stays clear as it compounds") are not in this PR.

A blind disposition bake-off — 3 versions × 3 dilemma tasks × 5 replicates, Opus, blind analyst panel classifying what each version did rather than scoring quality — found:

• The commit's structural rules are what drive completeness. On a completeness dilemma where the complete fix is subtractive (delete a duplicated source of truth), versions with the structural rules ("derive state instead of mirroring it," "keep one owner," "Replace; do not layer") completed 5/5; a version that had the "complete fix / cuts both ways" exhortation but weaker structural rules under-fixed 4/5. A vague "fix it fully" does not substitute for enforceable structure.
• Adding the counterweight on top of the structural rules was behaviorally neutral (a near-wash across the dilemmas).

So whether to also include the counterweight is a doctrinal call, not a behavioral one — parked on branch restore-excellence-counterweight. The structural additions in this PR carry no measured behavioral cost.

🤖 Generated with Claude Code