Fix storage, import maps, builds, and CLI handling

[ije]

Concurrent operations and failed I/O could corrupt cached files, lose import-map entries, or overwrite HTML after failed package resolution. This fixes those paths and several build, S3, and CLI errors found in the same review.

• Publish filesystem writes and lexer downloads only after successful completion; preserve existing files on failure and close S3 response bodies reliably.
• Follow S3 listing pagination, batch deletions with checksums, and surface per-object failures.
• Resolve the longest import-map prefix and synchronize dependency traversal, scope creation, CLI results, and web dependency loading.
• Preserve build arguments in declaration URLs and queue keys, reject corrupt splitting caches, and fix generated package metadata and CSS escaping.
• Repair npm bootstrap downloads, platform names, release tags, executable permissions, and exit status handling; preserve HTML and unmanaged imports during CLI updates.
• Remove dead code and unnecessary dependency traversal, replace version sorting with a linear scan, and avoid repeated directory-result copies.

Validation:

• go test -race ./...
• Staticcheck and git diff --check
• Debug lexer installation tests
• bun test cli/npm/bootstrap.test.mjs — 7 tests
• Deno transform, declaration, and raw-file suites — 6 tests against an isolated local server
• Version-resolution microbenchmark with 10,000 releases: median 3.44 ms → 2.50 ms across three runs on an Apple M4 Pro

S3 regressions use controlled HTTP responses; the credentialed live-S3 suite and full Deno suite are left to CI.