A C# member or parameter named with a verbatim identifier (`@class`, `@namespace`, `@event`, `@operator`) had its `@` dropped and was then read as the keyword, which opened fake scopes and changed member kinds. Verbatim identifiers are now kept as names and never matched as keywords or modifiers.
An Android manifest flag stored as an integer (`exported` = 0, `debuggable` = 1) was read as not set, so a component marked not exported could be reported as exported. Integer-typed values are now read as booleans as Android does, and only the un-namespaced `package` attribute names the package.
A GraphQL document that also declared a type-system definition without braces (`scalar`, `union`, `directive`, `extend`) or began with a description lost the operations after it, because the reader looked for a brace that never came. Those definitions and descriptions are now stepped over by their own grammar.
A reference into the middle of a string skipped that many characters, not bytes, so pseudo-C showed the wrong tail for wide strings and for text with multi-byte characters. The tail is now taken by byte offset (two bytes a unit for UTF-16LE), and a reference that lands inside a character is left as an address.
An imported function named with a symbol version (`memcpy@GLIBC_2.14`), a stub suffix (`@plt`) or a Darwin variant (`$UNIX2003`) was not found in the argument-count table, so pseudo-C dropped arguments set in earlier blocks. Those suffixes are now stripped before the lookup; stdcall `@bytes` names work as before.
When setup edits another app's YAML config, a key written with a space before its colon (`mcp_servers :`) was not found, so a second block was appended, and a plain key that merely starts with the same text (`craft-anything:dev:`) was taken for ours. Keys are now matched across optional whitespace before the colon and need a real key boundary after it.
Tracing a generated location through a source map made of sections found the original position but read the source snippet, the ignore flag and the source count from the outer index, which has none. So the snippet was missing, "ignored" was always false and the count was zero. They are now read from the section that covers the location, and sources are counted across sections.
* Use shared packet capture and per-frame protocol inspection
* Pin PCAP inspection to merged shared library
* Test shared PCAP truncation and public adapter parity
* Use shared WebAssembly inspection with app error compatibility
* Cover WASM resource indexes and component partial reports against oracle
* deps: pin craft-libs to the merged shared revision
Finishes #2037 for the last four page tools: bundle analysis, page watch, WebMCP tool discovery and the screenshot. They now answer a page's own dialogs the same way and list them in their results and summaries. The answering connect enables the Page domain itself, so each caller no longer has to.
A look at one game's folder or archive could finish after the person had picked another, and its list, profile or error then showed under the new pick. A finished look now carries the path it was asked for, and the Game files screen ignores a result for a folder or archive it no longer shows, leaving the new look running.
The four page tools that still attached without answering a page dialog (web_page_js_analyze, web_page_watch, web_mcp_tools_list and web_page_screenshot_take) could wait to their deadline when a page opened an alert or confirm. The shared answering connect now enables Page, which is what delivers Page.javascriptDialogOpening, and all four use it; their results and summaries list the dialogs. The bundle-analysis and screenshot results now carry the dialogs capture_page already collects, and WebMCP discovery drains its events after the probe.
Regressions: a fake CDP server opens a confirm on Page.enable and records Page.handleJavaScriptDialog; all four tools answer and report it. The existing seven dialog tests and the full browser module pass.
web_page_js_analyze, web_page_watch, web_mcp_tools_list and web_page_screenshot_take still attach without answering a page-opened dialog. Record the pending change before fixing it.
A look at one game's folder or archive could land after the person picked another one, so the old profile, archive list or file rows were shown under the new pick, and a save could name an item from the old game. The button job now carries its request, and game_files::tool_done ignores an Ok or Err result whose path no longer matches the shown folder (game_archives_list/core_target_triage) or archive (game_pak_list/game_unity_list/game_iostore_list/game_unity_bundle_list). Export toasts still say a file was saved.
Regressions: an old game folder's archive list/profile and an old archive's rows are ignored and leave the new look running; an old folder's error isn't shown, while the shown folder's error is. All three failed before the guard.
Follow-up to #2011, which answered a page's own dialogs only for the website session. The tools that connect to a page on their own (page show, walkthrough, coverage, listeners, module trace and the Electron page tools) now answer them the same way, so a page's alert no longer holds a call until its deadline: an alert or leave-page question is accepted and a confirm or prompt dismissed. Each result lists the dialogs and how they were answered, with declared secrets redacted, and the reporting code is shared in one place.
The tools that attach to a page themselves (web_page_show, web_walkthrough_record, the web-runtime coverage/listener/module tools, and the Electron page and walkthrough tools) connected without answering a page-opened alert or confirm, so the page's scripts stopped and the call waited to its deadline. They now answer dialogs as web_site_open does (alerts and leave-page questions accepted, confirm and prompt dismissed), enable the Page domain that delivers the event, and report each dialog's kind, first line and answer in the result and summary. Declared secrets in a dialog message are redacted. The shared dialog extraction moved to web_common so the website open and these tools use one reporter.
Regressions: a fake CDP server opens a confirm only after Page.enable and records Page.handleJavaScriptDialog; web_page_show, web_walkthrough_record, web_coverage_measure, web_listeners_list and web_modules_trace answer and report it. A unit test covers extraction, summary, empty/null and redaction.
Follow-up to #1965. build_file_edit read the whole file into memory before editing, so an oversized generated file could freeze or kill the app. It now uses the same 16 MiB capped read as build_file_read: a larger file is refused with a plain error and left unchanged, and text that is not valid UTF-8 is still refused.
The French catalog was missing 27 strings from recent work (the update checker, updates managed by a launcher or package manager, agent management and website sign-in), so the French UI fell back to English on them. Placeholders are unchanged.
A builder's edit read the whole project file into memory before replacing text, so a large generated file under app/ could freeze or kill the app. It now uses the same 16 MiB cap as a read, refusing a larger file before it is loaded and leaving it unchanged. A file at the cap still edits when the result fits the 8 MB write cap.
Regressions: an edit over the cap returns OverLimit and leaves the file unchanged; it failed before the fix with the write cap's BadInput after the whole file was read. A file exactly at the cap still edits down to the write cap.