There was an error while loading. Please reload this page.
📝 docs: state the hydration contract literally "Everything else travels" was not literal, and the gap is the dangerous direction: a null, a delegate, and a value that fails to serialize are all left out, and a reader could take that for protection. It is a robustness rule so one bad field cannot empty the payload — never a reason to believe a secret stayed behind. Also moves the System-exclusion rationale into the same paragraph as the rule, where someone deciding what to load will actually read it.
📝 docs: a ServerAction may USE a secret, never return one The previous wording said to resolve a secret inside a ServerAction "which returns only its result" — and an action's return value is serialized into the JSON response and delivered to the browser exactly as a field is. Read literally, the advice told the reader to return the token. Says it properly now: the action may use the secret server-side, and returns the ANSWER, never the secret that produced it. Same rule on both sides — what crosses is what the page may show. Caught in review of the change that introduced it. EN and pt-BR together.
📝 docs: every field of a prefetched page is public The payload is written into the served HTML, so a value a prefetch stores is readable by anyone who views the page source. Nothing said so, and the question came from someone who noticed a password field riding along — empty at SSR, harmless today, and exactly the shape that is not harmless the day a page loads a token. States the rule, the guidance (load what the page displays; a secret belongs inside a ServerAction), and the one exception: a dependency is skipped because the client resolves it. Also says why the exception is not "skip every interface" — IReadOnlyList<T> is how a component receives its items. EN and pt-BR in the same commit.
📝 docs: hreflang, and a mark that named the wrong release ServerIntegration gains the hreflang section (the policy, the two ready-made URL shapes, the three rules enforced for the app, and the DI caveat that cost an afternoon), Localization points at it from where someone would look, and the SeoBuilder table lists Alternate/AlternateDefault. Icons said gradients shipped in preview.30. The tag points at b57575f and the gradient commit came after it, so both languages now say preview.31.
📝 docs: the audit's findings — a false fence, two missing sections, one spelling A pass over the whole wiki, mechanical rather than read: every version mark checked against the tag that actually contains the commit introducing its symbol, every canonical page checked for a twin, every backticked identifier checked against the source. - Icons said gradients were DROPPED from the `.svg` subset. They are understood since preview.30, in both units, with `href` for a shared palette, and each target paints the run the way it natively can. The paragraph said the opposite in both languages, which is the exact shape of rot this wiki is prone to: the code moved and the page stayed confident. - The pt-BR Home lost its Development Guide section and the pt-BR Localization page never got "Photon windows adopt the machine's locale at launch" — the only mark missing from a twin. - Eight marks in five pt-BR pages said *Since* where the page teaching the convention says *Desde*. The site accepts both; a reader on GitHub sees the prose. Clean in the audit: 147 marks, none naming the wrong release; 36 canonical pages, none without a twin and no orphan twins; no backticked identifier naming an API the source does not have.
📝 docs: the visual editor page catches up with the editor The page was written when there was a preview and an identity mechanism and nothing else, and it has been saying "no editing from the canvas" through four slices that added exactly that. Nothing compiles a wiki, so a stale page stays green forever and the site publishes it. Now documented, EN and pt-BR: the three tiers a selected node is classified into and why they exist (a row built five hundred times in a foreach has one source site and no separate existence, so "delete this row" is wrong for it); the inspector, its three honesty rules — compiled before offered, WorkspaceEdit rather than a file write, never rewriting the form the file is written in — and the one level it descends so that padding and background are reachable at all; and the three structural gestures with the fence they live inside. The distinction the whole design turns on is now stated rather than implied: UNDERSTANDING is not fenced, only WRITING is. The editor opens code that already exists, and real code splits itself across helper methods and files. Both Roadmap pages carried a premise the work disproved on day one — that click-to-select would ride the V3 source maps eqc already emits. It cannot: the whole Build body is emitted through one call, so the finest position a map can name is the start of the method. The repo's ROADMAP said so; the wiki went on telling readers the opposite.
docs: the interaction contract catches up with the a11y pass — both languages
docs: the provenance header, in both languages
📝 docs: ServerIntegration speaks pt-BR