Record session-environment lessons and separate review axes

[geldmacher]

Context

This follows a review of Matt Pocock's skills v1.3.1 (mattpocock/skills at 24fe0ef7737efae15c87225755e9f6f5965e4888), announced in https://x.com/mattpocockuk/status/2107062474914550098. The sources read in detail were /retro and code-review.

Two gaps were worth taking. The rest was left out on purpose:

• /implement-spec, and the ticket-graph implementer flow around it, conflicts with Auto-Work and the written plan handoff. Planning stays the high-effort step. Implementation follows that plan. Host configuration chooses models; this plugin does not hardcode them.
• GLOSSARY.md and grill-me / grill-with-docs would add a required glossary and an interview. That fights the lean handoff, and a clear assignment already needs no interview.
• No new /retro command, no required CODING_STANDARDS.md, no parallel review subagents, and no Fowler smell list. Automated checks were already on the correction ladder from Prefer proportionate prevention when corrections recur #7.

What changed

• Learning (references/learning-work.md): the session or its log is a source for lessons about the agent's environment. The added cases are navigation, wasteful tool use, and steering instructions that do not change behavior. A mechanical lesson becomes a check. A judgment lesson becomes a standard the reviewer applies, outside implementation context. The existing order (structure, then CI, then bounded guidance, then human review) stays. A few nearby sentences were tightened so the addition fits the frozen context limits.
• Review (skills/review-work/SKILL.md): findings stay on two axes in the same read-only pass: deviation from the plan or spec, and repository standards or simplicity. On a simplicity question, use Efficiency's guidance when that plugin is already available.
• Credit is in docs/methodology-sources.md, with an Unreleased changelog note. No version bump.

Walkthrough rows in docs/behavior-validation.md, plus short notes in the working guide and project-improvement page, match those decisions.

Why

The correction ladder already preferred a structural fix or a check over another prose rule. What it did not say is to read the session itself for how the agent moves through the repository, or to keep judgment standards on the reviewer so implementation context stays lean. Splitting review findings the same way stops a standards pass from hiding a missed requirement, and the other way around, without a second reviewer.