v0.26.0 — the studio as an app, from an empty folder to the report

The studio becomes an app: from an empty folder to the report of what was built, step by step, without the commands. And a clean slate: everything a brief-to-app run left on the list, the test suite green on Windows.

Added

• The studio is an app you use without the commands (REQ-204): specgate studio opens on a route of seven steps — Project, Brief, Drafts, Questions, Review, Build, Report — each done, current or locked as the project's files say, with one next action (REQ-212). A simple mode hides the commands and an advanced one shows each action's; Spanish and English. Every write is a POST with the page's token; the overview it replaced is at /overview.
• A project from an empty folder (REQ-213): four questions — name, web/API/mobile, stack, agent — and the studio runs specgate init, writes the harness configuration and the agent's profiles, a brief to fill and a first commit. A folder with files is left alone, with the command that adopts it.
• The brief as a form (REQ-214): product, who uses it, and per module its rules, examples and open questions, saved into docs/brief.md keeping every other line; markdown in advanced mode; the brief rules come back with each save. A client's text can be pasted for the agent to structure (REQ-215).
• Drafting each module from the studio (REQ-216): in the background, one at a time, with its progress and the checklist in plain words, and a way to stop it.
• Review and approve a draft (REQ-206, REQ-207): specgate change approve <id> records that a person approved it, refused while the checklist has errors or a question is open; specgate change note <id> "<text>" --req REQ-x asks for a change in your words. Review in the studio does both.
• Build from the studio (REQ-208, REQ-209): what is ready, an estimate, the harness started in the background, its steps in plain words, and a stop that takes effect after the requirement in progress. Answering a question releases the requirements it held (REQ-210).
• The agent chosen per run (REQ-211): harness run --profile <name> and propose --from-brief --profile <name> run that profile of .harness/profiles.yaml; the studio lists the profiles with their cost hint and whether they stream usage.
• The report (REQ-217): what was built of the total, each requirement with the scenario that proves it, what is missing and why, the decisions taken, cost and time; downloadable as one HTML page.
• Answer a question from the live view (REQ-201): under each question the agent asked, Needs attention has a form to answer it and then mark it reviewed. The write, POST /live/answer, goes through the same code as specgate answer and needs the token its own page carries, a JSON body and a loopback origin; any other method than GET or HEAD gets 405.
• Every open question in the live view (REQ-203): an Open questions section lists what specgate questions lists, each with the same form, so drafts' questions are answered from studio too; a question the agent asked shows as waiting for an answer, not as a failure.
• A requirement used on a screen needs an end-to-end test (REQ-199): with e2e_dir in harness.config.yaml, a requirement whose actor uses web, mobile, tablet or desktop passes the gate only once a file there names it, and its prompt says so.
• A requirement an open change modifies is reworked (REQ-198): the harness gives it NEEDS_REWORK instead of NEEDS_STATUS_UPDATE, and its prompt says the tests and code are for the old text.

Changed

• The harness template shows the streaming agent (REQ-200): the Claude Code command that feeds studio's live view, and e2e_dir.
• Opening a module says which drafts stop growing (REQ-197).
• Status names the module of the next use case (REQ-193), and DRAFTING.md gives the draft's real limit with how to grow it (REQ-194).

Fixed

• The dry run names the agent only when one is set, so harness prompt prints no Agent none.
• The Cucumber gate on Windows (REQ-196): the message stream went to a path Cucumber did not write, so a filter matching nothing passed. The four tests that failed only on Windows pass, tests remove their temp folders, and docs:terminal / docs:journey say they need a POSIX shell.
• Ids reserved from a mention (REQ-202): "GIVEN REQ-999 is defined nowhere" in a scenario made the next change start at REQ-1000; only headings, matrix rows and req_range count now.
• No inline code split across changelog lines (REQ-195), checked by a test.