fix(#1461): loader never crashes on a malformed overlay (skip-with-warning); bound the source fetch

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.changeset/fix-1461-overlay-crash-dos.md

@@ -0,0 +1,7 @@
+---
+type: Fixed
+pr: 1475
+---
+**The capability loader never crashes on a single malformed overlay, and untrusted manifest/tar reads are size-bounded (ADR-1244 D2 invariant)** — `loadRegistry` now makes the WHOLE per-candidate overlay-processing body total: ANY throw from ANY validator or step (including the committed `validateCapability`, which dereferences a malformed array entry such as `gates: [null]` / `steps: [null]` / `contributions: [null]` before its shape check) drops just that one overlay with a skip-warning instead of escaping the loader, while `gatePointsOf` is hardened to be total over null/non-array/malformed gates. The final canonical `buildRegistry` compose stays guarded: a throw on the merged set falls back to the frozen first-party registry plus a warning, records each dropped gate-declaring overlay's declared gate as blocked (`incompatibleGateCapIds` / `blockedGates`) so a dropped blocking gate FAILS CLOSED, AND now clears `_overlay.commandRoots` in the fallback so no dropped overlay retains a stale command root. Previously a malformed-array throw or a compose throw escaped the loader and crashed every consumer (loop-resolver, config-loader, surface, capability-state, gsd-tools). On the source side, the capability resolver/staging now reads every untrusted `capability.json` (tarball / npm / git / local) via the shared bounded fd-reader (regular-file + 8 MiB cap) instead of a raw `fs.readFileSync`, so an oversized or FIFO/non-regular extracted-or-local manifest can no longer OOM or hang the resolver; the fetch (`realHttpsGet`) bounds the downloaded response to 64 MiB; and `stageValidated` now enforces ONE uniform aggregate byte-budget (`MAX_STAGED_BUNDLE_BYTES`, 128 MiB) over the STAGED bundle directory via a bounded streaming walk (cumulative byte + entry counters; symlink / non-regular entries rejected) at the common staging chokepoint — AFTER staging and BEFORE validation/promotion — so a huge source tree, git repo, npm package, or gzip/tar bomb is rejected (and its staging dir cleaned up) before promotion, uniformly bounding the RESULT of `copyDirRecursive` / `git clone` / `npm pack` / `tar -x` that were previously only timeout-bounded. `copyDirRecursive` itself is now STREAMING and BUDGETED: it enumerates each directory via `fs.opendirSync` + `dir.readSync()` (one entry at a time) and threads CUMULATIVE entry (`MAX_STAGED_BUNDLE_ENTRIES`, 100k) and byte (`MAX_STAGED_BUNDLE_BYTES`) counters through the recursion, failing closed the MOMENT either cap is exceeded DURING the copy — closing a residual where the former `fs.readdirSync(src, { withFileTypes: true })` materialized the ENTIRE directory-entry array into memory at staging time (BEFORE the post-copy budget walk), so a hostile source whose tree held a directory of millions of tiny files (fetch < 64 MiB, but a colossal dirent array) could OOM the process during the copy before the budget could fail closed; the post-copy walk is retained as a cheap belt-and-suspenders re-verification on what actually landed in staging. The spoofable per-member `tar`-header size parse (`parseTarMemberSize`, which mis-anchored on BSD `tar -tv` owner/group columns such as a `Jan` group → fail-open) was REMOVED in favor of that non-spoofable staged-dir budget; `assertSafeTarMembers` keeps its unambiguous traversal / symlink / hardlink NAME and TYPE guards. (#1461)
+
+<!-- docs-exempt: internal robustness hardening of the runtime capability loader/source seams (capability-loader.cts / capability-source.cts) — no user-visible API change; the never-crash + bounded-fetch guarantees are ADR-1244 D2 internal invariants -->