feat(prose): the conversation runs at product altitude — a raise is composed from the product, never the code

[leeovery]

What

A live research session (Fumi, platform-support, v0.7.27) raised a deep-dive finding as a compressed read-out of the report's code — four source snippets, a position at implementation depth, zero question marks — and the user could not tell what was being asked. Retold as a product scenario with three end states and one question, it was answered in a sentence. The same shape had already lost the user once earlier in the same session.

The altitude rule existed (product-lens.md) but was loaded at two sites only — investigation sign-off and review presentation — never by the research/discussion walk, never by the phases, never by the framework. The walk's one clause ("product perspective first") sat inside a 4,300-word state-machine protocol, and its nested making-it-land.md load was never read.

Changes

• altitude.md — new shared reference, loaded by framework.md beside voice: product first, code as evidence beneath it, internal names translated, mechanism the implementer's; a fork put to the user is composed as product end states or it is not the user's.
• composing-a-raise.md — the walk's raise composition extracted from background-agent-surfacing.md §G into its own reference loaded at the point of composition (the move impl+review(prose): a surviving decision is raised, never rendered — the walks share one arm #1079 made for decisions): the problem retold upward from a report that may be written in code, the position at its earned firmness, and a closing beat that is a literal question only where the sides compose as product end states.
• Agents — deep-dive, research review, discussion review, and synthesis open every finding on the product; code beneath as evidence. The record stays technical.
• Phase guidelines — research and discussion carry their altitude line. product-lens.md composes with altitude instead of restating it.
• Case — research-raises-a-code-shaped-finding: a fixture-authored deep-dive written in code with a product fork the user owns hiding under three mechanisms. Every walk case until now used a product-shaped report, so the corpus could not see this drift. All 109 nets gain altitude.md; the 33 walk nets gain composing-a-raise.md.
• CLAUDE.md / CONVENTIONS.md updated for the framework list and the register.

Gate

npm test 2750/0, npm run typecheck clean. Walks not run — every case intersects (framework changed); the ones that exercise the change: research-raises-a-code-shaped-finding, discussion-raises-problem-then-position, research-review-walks-the-lanes, discussion-walks-an-early-review, discussion-resumes-review-walkthrough, discussion-promotes-a-batch-item.

🤖 Generated with Claude Code