docs: note that the book's snapshots aren't covered by task prqlc:pull-request

[prql-bot]

CLAUDE.md presents task prqlc:pull-request as "sufficient for most changes" before handing work back. It isn't sufficient for a change to prqlc's output: the task's package list is the prqlc crates and their bindings, and the book's tests live in mdbook-prql, which no task in either Taskfile runs. Three of those tests read prqlc's output — the book examples' compiled SQL (snapshots under web/book/tests/documentation/snapshots/), their RQ serialization (book::test_prql_examples_rq_serialize), and the dialect list that web/book/src/project/target.md documents (book::test_target_dialects_documented). A change to any of the three can pass the documented pre-return gate and still go red on test-rust.

That is not hypothetical — it happened on #6297 in this repo yesterday. task prqlc:pull-request and a full -p prqlc run were both green locally; test-rust and measure-code-cov then went red on documentation__book__reference__syntax__s-strings__3.snap (job), costing a CI round for a snapshot the local gate could have caught in seconds.

This adds the missing command next to the gate it belongs to, with the trigger stated broadly enough to cover all three tests rather than only the snapshot case that prompted it. No test accompanies it: the change is documentation, and the claim it corrects is about which packages a task enumerates, which the Taskfiles state directly.

[prql-bot]

Self-review. The mechanism the note describes holds — prqlc:pull-request enumerates prqlc-parser, prqlc, prqlc-macros, compile-files and the bindings, no task in either Taskfile reaches mdbook-prql, and the added command runs the package's 10 tests green.

The trigger is narrower than what the command guards, though, and both the prose and the code comment state it. mdbook-prql's tests also fail on prqlc changes that leave compiled SQL untouched: book::test_prql_examples_rq_serialize serializes every book example's RQ, so an RQ representation change breaks it, and book::test_target_dialects_documented cross-checks web/book/src/project/target.md against an exhaustive match over dialect support levels, so adding a dialect breaks it without moving a single .snap. Both go red on the same test-rust job for the same reason #6297 did — so an agent reading "when the change alters compiled SQL" literally skips the gate on exactly the changes it was added to catch. Suggestions inline widen it to prqlc's output generally; the prose one is wrapped to match proseWrap: always.