Summary

Adds a read-only, server-side manifest listing endpoint, implementing the read half of #5535.

GET /manifest/{address}/{prefix}

Bee currently has no way to enumerate a manifest's contents over HTTP — clients must download and walk the Mantaray trie themselves, chunk by chunk via /bytes (one round trip per node). This endpoint walks the trie on the node and returns entries as JSON, turning O(trie nodes) round trips into O(pages). Semantics mirror S3 ListObjectsV2:

Response: { entries[], commonPrefixes[], truncated, nextMarker }.

This is deliberately listing only. The companion server-side mutation endpoint sketched in #5535 is a separate follow-up.

Implementation

Everything needed already existed in the tree:

• pkg/manifest — a small optional EntryWalker interface (+ ErrStopWalk), implemented as WalkEntry on the mantaray manifest over the existing sorted WalkNode. This keeps the core manifest.Interface and simple.go untouched and doesn't expose the trie representation to callers.
• pkg/api/manifest.go — manifestListHandler, mirroring the /bzz download path's manifest resolution (loadsave.NewReadonly → NewDefaultManifestReference) and reusing the ACT decryption chain, so encrypted/ACT manifests resolve unchanged. Pagination uses a last-consumed-path marker that is correct for both recursive and delimiter modes.
• openapi/ — path + ManifestList/ManifestListEntry schemas.

No new storage, protocol, or incentive-layer behavior — purely a local read amplification of data the node already serves.

Testing

• Unit (pkg/api/manifest_test.go): recursive, delimiter=/, prefix, multi-page pagination round-trip, missing-prefix 404, and over-cap-limit 400, over a real mantaray fixture (incl. the empty root-metadata entry). go test ./pkg/api/ ./pkg/manifest/... green; go vet and gofmt clean.
• Live: verified end-to-end against a funded Bee 2.8.1 light node on mainnet, listing a manifest built by Bee's own POST /bzz uploader. All cases correct, and sizes=true returned exact byte lengths from real chunk spans. The live run caught one bug the synthetic test missed — the RootPath metadata entry carries a 32-zero-byte reference (not swarm.ZeroAddress, which has nil bytes), so it leaked as a / entry until the skip was aligned with the IterateAddresses convention.

Known limitations / scope (draft — feedback welcome on the API shape)

Opening as a draft to agree the API before polishing. Deliberately out of scope here, happy to address based on review:

1. Delimiter mode still descends full subtrees. WalkNode has no pruning hook, so delimiter=/ walks a subtree fully to emit one commonPrefixes entry — output is correct but not yet the cheap shallow listing it should be. The clean fix is a pruning-aware walk in the mantaray layer.
2. Each page re-walks from the start (after is a post-filter), so paginating N pages is ~O(N × subtree) node loads. A trie-seek to after would fix it.
3. prefix is directory-boundary-only, not an arbitrary S3 byte-prefix (e.g. prefix=da 404s; data/ works). Documented as such; could add HasPrefix-style handling.
4. No per-request node-load budget yet (the DoS mitigation noted in API: server-side manifest listing endpoint (S3-style), with optional manifest mutation as a follow-up #5535); sizes=true on encrypted refs is silently unsupported. The limit hard cap bounds buffered response memory but not the number of trie nodes walked to build a page.
5. Root-level manifest metadata (e.g. website-index-document) is currently dropped from the listing; API: server-side manifest listing endpoint (S3-style), with optional manifest mutation as a follow-up #5535 proposes surfacing it as a top-level manifestMetadata — easy to add if wanted.

Refs #5535.