perf(@angular/build): verify cache metadata before reading file and eliminate sqlite read-locks

[alan-agius4]

Summary

When using persistent caching in @angular/build, PersistentLoadResultCache.get() previously read the entire target file from disk via readFile() and computed a SHA-256 hash before querying the persistent L2 store.

This introduced two primary bottlenecks:

1. Redundant Disk I/O & Hashing: On cold cache misses, every project file was read and hashed by the cache, found to be absent in the store, and then read from disk a second time by esbuild. On warm cache hits, reading and hashing multi-megabyte JavaScript files from disk was redundant when a fast stat() check (mtimeMs + size) would confirm whether the cached entry remains valid.
2. SQLite Read-Lock Contention: In SqliteCacheStore, #queueAccessUpdate() triggered a synchronous BEGIN IMMEDIATE TRANSACTION; write transaction every 100 cache reads to update last_accessed. In multi-process worker pools and parallel builds, this converted concurrent read operations into serialized write transactions that blocked on SQLite busy timeouts.

Optimizations

• Metadata Verification Before Read: Cache keys are now derived from the global configuration hash and path, querying the persistent store before performing disk file reads.
• Watch Files & Target Metadata: Records target file metadata alongside dependency watch files in watchFilesMetadata.
• Fast-Path Metadata Validation: Validates cache hits using fast-path metadata comparison (mtimeMs and size) for both the target file and its dependencies, only reading content and hashing on disk if timestamps changed.
• Deferred Access Updates: Defers SQLite last_accessed timestamp updates via unref'd timer and flushes batch updates on store close, preventing write locks from blocking concurrent cache reads.

Benchmark Results

In benchmarks on 309 project files:

• Cold cache miss latency: dropped from 98.5 ms to 1.5 ms (65.2x faster).
• Warm cache hit latency: dropped from 77.7 ms to 17.9 ms (4.35x faster).
• Concurrent latency (4 worker processes): dropped from 143.6 ms to 17.5 ms (8.2x faster).

[gemini-code-assist]

Code Review

This pull request simplifies the cache key calculation in PersistentLoadResultCache by removing the file content from the key generation, which avoids redundant disk reads when checking the L2 cache. Additionally, it adjusts the flushing behavior in SqliteCacheStore by removing the immediate flush threshold and increasing the timeout. The reviewer suggested simplifying and unifying the deduplication of watch files using a single Set initialization to prevent redundant operations even when filePath is undefined.

[alan-agius4]

This PR was merged into the repository. The changes were merged into the following branches:

• main: d11a663