emit build-database for IDEs: one failing member loses the whole database, and build programs cannot tell a plan pass

[Sunrisepeak]

mcpp-language-server gets a project's model from mcpp emit build-database --format json. Reproducing Sunrisepeak/mcpp-language-server#23 on a real 5-member Windows workspace (GalTranslPP: vcpkg + Qt 6, mcpp 2026.9.25.1, windows-2025) showed two properties of emit that cost the editor its model. Both are behaviour the spec states today, so this is a request, not a bug report. Tracked on our side as UP-M2 / UP-M3 in Sunrisepeak/mcpp-language-server#24.

1. One member's failure loses every set (SPEC-005 R5.2)

• GPPGUI's build program calls mcpp::dep_bin("gpp.updater", "Updater"). Building that host tool ran its own lupdate check action, which failed on that machine (lupdate.exe could not load Qt6Qml.dll).
• Result: MCPP_BUILD_DATABASE_PLAN_FAILED: GPPGUI: building host tool 'updater:Updater' failed, and no data — the four members that planned fine got no database either.
• Effect: the language server falls back to a guessed model for the whole workspace (a toolchain it found, no include directories); clangd then crashed repeatedly on it.

Ask: when some members plan and others fail, return data with the sets that planned, plus an error diagnostic per failed member, under a code a consumer can tell apart from a total failure (e.g. MCPP_BUILD_DATABASE_PARTIAL).

2. A build program cannot tell it is running for a plan (SPEC-005 R2.5)

• emit runs build.mcpp exactly as build does and builds dependencies' host tools, with no signal in the MCPP_* contract.
• A build program that checks the full environment (a Qt root, a vcpkg .lib per library) — or a host tool whose build runs project checks, as above — fails emit on every machine that cannot build yet. That is exactly when an editor most needs the model.
• mcpp::action edges already do not run under emit (R2.2); the problem is work done in the program itself and in host-tool builds.

Ask (any of):

• a documented signal, e.g. MCPP_PLAN_ONLY=1 / mcpp::plan_only(), so a program can skip checks and heavy work while planning;
• under emit, a build-program failure becomes a diagnostic for that member instead of failing the command (pairs with 1);
• under emit, do not run the check actions a host-tool build declares.

Evidence

• Sunrisepeak/GalTranslPP PR feat: compile .c sources via a separate C rule #1 (temporary CI): run 36164741493 fails as above (Qt without qtdeclarative); run 36178100611 plans all 227 units once Qt is complete.
• Related: feat: mcpp emit build-database —— 不写工程目录的 S1 构建数据库（模块图、标准库模块、监视输入） #636, emit build-database 供 IDE 调用：离线缺依赖的错误码、规划期间不泄漏调用方的管道、xlings 调用要有期限（附联网效应、auto_refresh、索引镜像） #648.

[speak-agent]

Released in mcpp 2026.9.26.2 (#702, squash c109fdd6).

mcpp emit build-database now plans every selected member on its own:

• A member whose planning fails contributes no sets and one error diagnostic whose path names its mcpp.toml. data is present when any member planned, and its watch also lists each failed member's mcpp.toml and build.mcpp. The exit status is 1 whenever an error is present.
• Under emit, a host tool that does not build is the warning MCPP_BUILD_DATABASE_HOST_TOOL_UNBUILT. Planning continues, and the requesting build program receives the path the tool would be published at. mcpp build still fails as before.
• A package whose build program fails is described without that program's directives, with the error MCPP_BUILD_DATABASE_PROGRAM_FAILED whose path names its build.mcpp.
• The plan-only signal is not added: verification belongs in check actions, which emit does not run (design record §4.7, SPEC-007 R1.2, R1.4).

S2 0.3.0 states the partial answer (S2-3.4-12, S2-3.4-13). It landed in Sunrisepeak/mcpp-language-server#26 together with the consumer side, and mcppls 0.0.5 reports each part the producer could not describe as producer-partial.

The same release also changes the compile database, based on the report that accompanied this issue:

• one database per configuration;
• directory is the output directory;
• a module interface's entry carries its language flag;
• the standard-library units are listed.

Readings:

• e2e 787, 788 and 789 fail on 2026.9.26.1 and pass on 2026.9.26.2.
• In an xlings sandbox, with CN mirrors and mcpp installed from the published index, the release's own e2e scripts pass against the published binary: 20 ok, 0 failed. The same script against 2026.9.26.1 fails exactly the 16 checks this release changes.
• CI of 2026.9.26.2: one compile database per configuration, emit plans every member, prepare actions and runtime search directories, DLLs beside a Windows program, link flags as words #702 is green on every platform except the two xcode-27 legs of ci-macos xcode-27: ld64.lld cannot parse arm64e.x1 in either available SDK (upstream, tracked) #669, which fail the same way on main.