fix(progress) + feat(insights): eight defects, and a conclusion you can remove

[mrviduus]

Two things that arrived together: the progress-path debt recorded while tracing the assistant handoff, and what a three-way consilium on insight categories turned up while arguing.

Progress path — four defects fixed, two deliberately not

"Mark as finished" wrote two different things. Web sent the {"type":"end"} sentinel; mobile, from the same menu on the same kind of book, sent scroll:<lastSlug>:0 — so a book marked finished on the phone reopened at the top of its last chapter. One definition now (packages/shared/src/reader/progressLocators.ts), both clients import it.

The first write for a book skipped the rules. PUT /me/progress/{editionId} had two code paths writing one row: an update that checks ProgressUnit.IsTrusted, and an insert that assigned request.Percent by hand. So the first write for an edition kept an untrusted number every later write would have refused — and a book finished in a single write never got a CompletedAt. The insert now hands the row to the same ApplyProgressUpdate. ProgressUnitEndpointTests fails against the previous build and passes against this one.

ReadingProgressDto existed twice in TypeScript with nothing linking the copies.

A tap that did nothing now says so — "Talk this book over" swallowed openURL's rejection, which is right for "you have no browser" and indistinguishable from a broken button.

Not fixed, on purpose (written up in assistant-handoff.md so nobody "fixes" them): LocatorKind on the catalog path would refuse the very sentinel it was meant to protect — catalog editions have no second coordinate space; and markAsUnread not lowering MaxChapterNumber is moot, because that column's only reader was the RAG gate deleted with ask_book.

The web insights section threw on every signed-in book page

Found by the consilium, not by a test. The web-local insights client unwrapped .items — an envelope GET /me/insights does not send (it returns a bare list; the envelope belongs to /me/mcp/keys, where the line was copied from). The component stored undefined and read .length off it on the next render. Every signed-in BookDetailPage and UserBookDetailPage, with or without insights, in the change that was meant to make that section render at all. Its own test mocks the module wholesale, so it could not see the shape; the new test pins the wire shape instead.

A conclusion can be removed

Saving again replaces the insight for the same chapter, so a poor conclusion about the right chapter self-corrects. One filed against the wrong chapter never does — nothing revisits that slot. It was permanent.

DELETE /me/insights/{id}, reader-only, 404 (not 403) for someone else's id, hard delete. Deliberately not an MCP tool: save_insight's worst case is one bad paragraph removed in a tap; a delete tool's worst case is a year of конспект gone, from a stateless bridge that cannot confirm intent.

Also

Five web modules and one stylesheet with no importers deleted — three were listed as live in CLAUDE.md's own hook inventory.

Verification

• dotnet format textstack.sln --verify-no-changes — clean
• dotnet test tests/TextStack.UnitTests; integration --filter Insights (18) and --filter ProgressUnit (2) — pass against a locally rebuilt API
• web 709 + 2 new delete-path tests, mobile 369, shared 449 — pass
• tsc --noEmit on web and mobile — clean

🤖 Generated with Claude Code

https://claude.ai/code/session_011rgEMvYYi4Egj99dVtvm6E