feat: Incremental Build (reorganized commits)

.claude/skills/incremental-build/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: incremental-build
+description: >
+  Work on the incremental (delta) build system in @ui5/project: build caching,
+  resource indexing, hash trees, stage caching, build server, file watching,
+  task execution caching, and delta builds.
+when_to_use: >
+  TRIGGER when: the user asks about or wants to modify code related to incremental builds,
+  build caching, resource indexing, hash trees, stage caching, build server, file watching,
+  task execution caching, delta builds, resource tags in build context, or any component
+  listed in the Component Map in architecture.md.
+  DO NOT TRIGGER when: the user is working on unrelated CLI commands, general tooling,
+  or non-build features.
+user-invocable: true
+---
+
+# Incremental Build Skill
+
+You are working on the incremental (delta) build system in `@ui5/project`. Read `architecture.md` in this skill directory for the full architecture reference, including the component map, key flows, caching architecture, and data structures. Read `performance-investigation.md` for guidance on profiling builds, reading perf logs, and known performance peculiarities.
+
+## Guidelines for Working on This Code
+
+1. **Always read the source file before modifying it.** The Component Map in `architecture.md` tells you where each piece lives.
+2. **Understand the cache flow direction.** Changes propagate: source change -> index update -> signature change -> cache miss -> task re-execution.
+3. **Be careful with tag side effects.** `getTags()` is not a pure read -- it triggers lazy tag application. Avoid calling it in unexpected contexts.
+4. **Respect abort signals.** Any long-running operation should check `signal?.throwIfAborted()` periodically.
+5. **Test with incremental rebuilds.** A single build passing is not enough; the interesting bugs appear on the second and third builds after file changes.
+6. **Watch for stale stage readers after abort.** If a build is aborted, stage writers may contain partial results that shadow source files.
+7. **Signature stability matters.** Any change to how hashes are computed (e.g., adding new fields to hash input) invalidates all existing caches.
+8. **SharedHashTree operations must go through TreeRegistry.** Never mutate a SharedHashTree directly; always schedule via the registry and flush.
+
+## Known Constraints
+
+- `StageCache` (in-memory) has no `clear()` method -- stage entries persist for the lifetime of the `ProjectBuildCache` instance
+- `ProjectBuildContext` instances (including their caches) are reused across sequential builds of the same project within a session
+- `resource.getTags()` has side effects: it triggers `#applyCachedResourceTags()` and creates `MonitoredResourceTagCollection` instances, which modify tag collection state. This means calling `getTags()` during hash tree operations (e.g., in `TreeRegistry.flush()`) can affect the stage pipeline state.
+- `updateProjectIndices` uses `project.getReader()` (stage pipeline reader) to read resources. After an aborted build, stage writers may contain in-memory resources that shadow updated source content, causing stale reads.
+- `getSourcePaths()` and `getVirtualPath()` are NOT consistent with `getSourceReader()` for all project types. Module has no `getVirtualPath()` (base class throws). ThemeLibrary's `getSourcePaths()` returns only `[src/]` and `getVirtualPath()` only maps `src/`, but `_getReader()` also includes `test/` when `_testPathExists`. Any optimization that replaces `sourceReader.byGlob()` with direct filesystem enumeration via `getSourcePaths()` + `getVirtualPath()` must handle these mismatches.

.claude/skills/ui5-fs/SKILL.md

@@ -0,0 +1,67 @@
+---
+name: ui5-fs
+description: >
+  Work on the @ui5/fs package: the virtual file system abstraction layer providing
+  Resource, Adapters (FileSystem, Memory), Reader Collections (ReaderCollection,
+  ReaderCollectionPrioritized, DuplexCollection, WriterCollection), specialized readers
+  (Filter, Link, Proxy), resource tagging, monitoring, and the resourceFactory API.
+when_to_use: >
+  TRIGGER when: the user asks about or wants to modify code related to @ui5/fs,
+  the virtual file system, resources, adapters, reader collections, writer collections,
+  DuplexCollection, resourceFactory, ResourceTagCollection, MonitoredReader, fsInterface,
+  ResourceFacade, or any file within packages/fs/.
+  DO NOT TRIGGER when: the user is working on builder tasks/processors, CLI commands,
+  server middleware, or project graph resolution that does not touch the FS layer.
+user-invocable: true
+---
+
+# @ui5/fs Skill
+
+You are working on the `@ui5/fs` package — the virtual file system abstraction layer for UI5 CLI. Read `architecture.md` in this skill directory for the full architecture reference, including the class hierarchy, API surface, adapter internals, and collection patterns.
+
+## Package Location
+
+Source: `packages/fs/lib/`
+Tests: `packages/fs/test/lib/`
+Fixtures: `packages/fs/test/fixtures/`
+
+## Guidelines for Working on This Code
+
+1. **Always read the source file before modifying it.** The Component Map in `architecture.md` tells you where each piece lives.
+2. **Understand the content type state machine.** Resources have internal content types (BUFFER, STREAM, FACTORY, DRAINED_STREAM, IN_TRANSFORMATION) with strict transitions. Never bypass `modifyStream()` for content transformations — it handles mutex locking and state management.
+3. **Respect the mutex.** Resource content access is protected by `async-mutex`. Concurrent `getBuffer()` / `getString()` calls are safe, but `modifyStream()` acquires an exclusive lock. Never hold a reference to content across an `await` that could trigger a transformation.
+4. **Virtual paths are POSIX-absolute.** All virtual paths must be absolute POSIX paths (start with `/`). Base paths must end with `/`. Adapters normalize patterns relative to their `virBasePath`.
+5. **Content parameters are mutually exclusive.** When creating a Resource, only one of `buffer`, `string`, `stream`, `createStream`, or `createBuffer` can be provided. The `createBuffer`/`createStream` factories enable lazy loading.
+6. **byGlob randomizes result order.** `AbstractReader.byGlob()` intentionally shuffles results to prevent consumers from relying on ordering. Do not assume or depend on glob result order.
+7. **Clone semantics matter.** Resources are cloned on retrieval from adapters. `resource.clone()` creates an independent copy including content. When modifying resources from collections, understand whether you're working on the original or a clone.
+8. **ResourceFacade is immutable.** `ResourceFacade.setPath()` throws. The facade wraps a resource with a different virtual path (used by Link reader). Use `getOriginalPath()` to get the underlying path.
+9. **Tag format is strict.** Tags follow the pattern `"namespace:Name"` — namespace is lowercase alphanumeric, name is PascalCase. Tags are validated on set. Use `ResourceTagCollection` or `MonitoredResourceTagCollection` for tag operations.
+10. **Test with both FileSystem and Memory adapters.** Behavior can differ between adapters (e.g., FileSystem uses `globby` while Memory uses `micromatch`; FileSystem has lazy content loading via factories, Memory clones on read).
+
+## Known Constraints
+
+- `Resource.getStream()` is deprecated — use `getStreamAsync()` or `getBuffer()` / `getString()` instead. The deprecation warning is only logged once per process.
+- `Resource.getStatInfo()` is deprecated — use `getLastModified()`, `getSize()`, `getInode()` instead.
+- FileSystem adapter's `write()` uses `fs.copyFile` for unmodified resources (optimization). It also detects same-source-same-target writes and skips them, but switches to buffer-based writes if the content was modified (to avoid stream read-during-write conflicts).
+- Memory adapter auto-creates virtual directory entries in its hierarchy on `write()`.
+- `WriterCollection` matches the **longest prefix** (greedy match) when routing writes to writers.
+- `DuplexCollection` uses an internal `ReaderCollectionPrioritized` with the writer first, so written resources shadow the reader's resources.
+- `fsInterface` provides a Node.js `fs`-compatible wrapper but only implements `readFile`, `stat`, and `readdir`. `mkdir` is a no-op.
+- `Resource.getIntegrity()` computes SHA-256 via `ssri` — this triggers full content loading if not already loaded.
+- Monitored readers/writers (`MonitoredReader`, `MonitoredReaderWriter`) prevent reads/writes after being sealed. They track all access patterns for build caching analysis.
+
+## Running Tests
+
+```bash
+# All tests
+npm run unit --workspace=@ui5/fs
+
+# Single file
+cd packages/fs && npx ava test/lib/Resource.js
+
+# Verbose with logging
+npm run unit-verbose --workspace=@ui5/fs
+
+# Coverage
+npm run coverage --workspace=@ui5/fs
+```