compile: prune modules reachable only through unused re-exports of side-effect-free packages (Bun's graph: 4,063 modules; perry compiles and initializes 7,897 for the same OpenCode entry) #10180

Parent: #10106 (OpenCode startup) / #10107. Related: #9133 (full-source builds), #10173 (mid-size literal compile cliff), the per-function code-size campaign.

Observation

For the same entry (packages/opencode/src/index.ts, OpenCode v1.18.30, bun condition) Bun's bundler graph has 4,063 modules; perry compile --platform bun collects and compiles 7,897 (audit: 2,067 first-party, 5,830 from 422 packages). The native binary then runs every one of those module initializers before main parses argv, which is most of the 2.0 s user CPU of opencode --version (bun: 0.4 s), and every module is ~130 KB of .text on average in the 1.05 GB binary.

Where the extra modules come from (from perry-cache/audit.json vs Bun's metafile, both saved under secret-tests/opencode-1.18.30-inventory/):

• Bun drops modules that are only reached through unused re-exports of side-effect-free packages ("sideEffects": false / pure barrels): remeda 24 modules in Bun vs 170 in perry, effect 225 vs 294, openai 145 vs 152, ai (227 in perry) similarly; barrel index.ts files that export * from "./x" pull in every sibling even when the importer names one symbol.
• 31 packages are present in several versions; @ai-sdk/provider-utils in 11 versions = 1,012 modules, @ai-sdk/provider 6 versions = 648, zod 3 versions = 189, semver 3, @ai-sdk/openai 3, entities 3. Bun's graph has 3 provider-utils versions. Perry follows each provider package's nested copy; that is correct resolution, but most of those copies are reached only through barrels of providers that OpenCode loads lazily via import() (BUNDLED_PROVIDERS in src/provider/provider.ts:113).
• node/bun condition twins are only 6 files, negligible.

Ask

Collection-time pruning that matches the bundler's semantics without changing program behaviour:

1. For packages that declare "sideEffects": false (or whose modules have no top-level effects by analysis), do not compile or initialize modules that are reached only through export { x } from / export * from re-exports whose bindings the importer never uses (bundler-style live-binding tree shaking at the module level; the module stays available for dynamic import).
2. Keep dynamic-import-only subgraphs out of the eager init order and out of the eager .text if possible (verify with the OpenCode graph that prettier/typescript/babel are not initialized on --version).
3. Report the pruned count in the compile summary (Found N module(s): …, M pruned as unreferenced side-effect-free re-exports).

Acceptance

• OpenCode v1.18.30 collects within ~10 % of Bun's 4,063 modules; opencode --version user CPU and binary .text reported before/after.
• A fixture with a sideEffects: false barrel proves the unused sibling is neither compiled nor initialized, and that a sibling with a top-level side effect (or a package without the flag) is still initialized in ESM order.
• Node-parity suites unchanged.