Bookmark editor: recorded implementation evidence

This is a dated evidence record for the runnable bookmark editor, not the lesson instructions. Start with that page to see the interface, run the app and follow the save/typing scenario.

Recorded 2026-09-22. Python 3.12.14, SQLite 3.53.1, Node 24.19.0.

Check	Evidence / scope
Python HTTP tests	5 passing methods against a real local HTTP server and temporary SQLite file; concurrent conditional writers, owner isolation, tied cursor, durable idempotent replay, validation
Browser build	`node build.mjs` succeeds; Node strips TypeScript syntax, which does not verify types
Strict TypeScript	Separate `tsc -p tsconfig.json` succeeds with strict checking, DOM/ES2022 types and no emit
Real browser	7 Chromium scenarios pass against actual HTTP + SQLite: A/B race, conflict/focus, old refetch, idempotent retry, ignored-abort disposal, search/empty/error/retry, and failed new search cannot reuse an old cursor
Query measurement	100,000 rows, 90% Alice, timestamp ties of ten, 50-row page, 15 warm queries; before median 19.855 ms, after 0.030 ms
Before query plan	`SCAN bookmarks`; `USE TEMP B-TREE FOR ORDER BY`
After query plan	`SEARCH bookmarks USING INDEX bookmarks_owner_created_id (owner_id=? AND (created_at,id)<(?,?))`

The query experiment checks equal results before/after the index. Timing varies by machine and cache; this is local query evidence, not measured browser interaction latency. Browser response fixtures use explicit promise gates. A saved screenshot was visually inspected: the server shows Edit A, input retains Edit B, and the status announces that the newer draft is unsaved. Tests regenerate this screenshot locally.

Independent review found and reproduced two browser/API defects before this evidence was finalized: failed search could reuse an old query's cursor, and whitespace-padded titles could pass validation but violate SQLite length constraints. Both were repaired with regressions. The query/cursor state resets together; API validation counts the exact title that is stored and rejects oversized version integers.

No AWS deployment, production login, full screen-reader evaluation, offline mutation queue or background enrichment is claimed. The real store is SQLite; API tests do not substitute an in-memory fake for it. Strict tsc checking requires an additional TypeScript installation; the check above used a separately installed compiler, without adding it as an application runtime dependency. Chromium used a packaged executable with system fonts in this execution environment; normal local Playwright installation instructions are in the README.