Let a sequence step show several code chunks

[jykim256]

Stacked on #627. Please merge #627 first. This branch sits on top of it, so the diff here shows only the new commit once #627 lands.

What changed

A sequence step could point at one code range. Real steps often touch two or three places, so the author had to pick one. Now a step can carry several.

• A step takes a new optional sources array, with 1 to 10 ranges.
• A step still needs exactly one of source, sources, explanation, or code. The error message now says so.
• When the guided tour lands on the step, it shows every chunk.
  • Chunks in the same file (same side and pins) share one code card. The card gets several ranges, so the exact lens from Keep tour steps and code peeks clipped to their pinned lines #627 folds the code between them.
  • Chunks in other files get their own cards, stacked in the order they first appear.
• The first chunk anchors the step. Jump-to-source on a card opens that card's first chunk.
• A step whose sources has only one entry looks exactly like a step with source.
• The MCP edit guidance and the scratchpad instructions now mention sources.

Authoring guidance

• instructions/authoring.md now tells authors to use sources when a step's story spans several places, in reading order.

Stored data

There's no migration. sources is additive. Stored steps that use source parse, render and tour exactly as before.

Each chunk becomes its own document reference, with the id <step id>:<index>. That means each chunk gets the same checks a single source gets: whitespace-only ranges get rejected, pins get validated, and staleness is tracked per chunk. A single-source step keeps its old reference id, so existing stale-source state still lines up.

How I checked it

• pnpm --filter @dev.fast/review typecheck and tsc --noEmit -p packages/review/app pass.
• pnpm lint passes. oxfmt --check passes on the changed files.
• pnpm --filter @dev.fast/review test:node: 1196 passed. One run hit a failure in src/sharing/export.test.ts (the diff-count watch stream test). It passed on four reruns of this branch and on the base branch, so it looks like a timing flake under load.
• Canvas browser suite (vitest --project browser): 46 files, 275 tests pass. That includes a new test where three chunks across two files make two editors, with the same-file editor getting both ranges.
• Canvas node suite (test:node, Node 24.18.0): 48 files, 247 tests pass, including the new sequence-tour.test.ts and code-peek-groups.test.ts.
• New server tests: a step with three chunks is stored, a step with both source and sources is rejected, and a whitespace-only range in sources[1] is rejected.

Things to look at

• CodePeekCard now wraps a new CodePeekFileCard that takes a list of sources. Single-source callers (code peeks, side peeks) get the same title, ranges and reveal behavior as before.
• The tour uses a new peek content kind, sources. Only sequence steps produce it right now. Flow-diagram attachments still make one tour stop per source. I left that alone on purpose, but it could adopt the same stack later.
• I picked 10 as the max. It's easy to change if that's too tight.
• No desktop (Code - OSS) changes. The desktop lens already handles several disjoint ranges in one file.

🤖 Generated with Claude Code

[sidkmenon]

hey! @jykim256 tysm for the contribution !! sorry for the delay on this, have been working on some diffr data model improvements to fix up the diagram code snippets.

i'm a bit confused by this though - "the multiple code snippets in a stop" doesn't feel very natural for a sequence diagram? are the other code snippets like context or something? and the limit of 10 feels a bit arbitrary.