fix: percent-encode storage resource paths on the wire (#124)

[andrii-novikov]

Applicable issues

• fixes files.move_to / copy_to 500 on paths with spaces or reserved characters (raw path sent in JSON body) #124

Description of changes

DIAL Core parses every storage resource path — both in the request URL (upload/download/delete/get_metadata) and in JSON body fields (sourceUrl/destinationUrl for move/copy, url for permission grants) — as a percent-encoded path (each segment is URL-decoded via UrlUtil.decodePath, which runs new URI(segment)).

A raw reserved character in a filename therefore broke things:

• # / ? were truncated by the client's urlparse in get_api_path before the request was even built → wrong path → 404.
• [ / ] / space and friends reached Core unencoded → new URI(...) throws URISyntaxException, which Core wraps in a bare RuntimeException → 500.

The URL-slot ops only worked by accident (httpx encodes the URL path on the wire); the body-slot ops (move_to/copy_to) had nothing encoding them at all.

Fix:

• Add percent_encode_resource_url() (stateless helper) and DialStorageResourceMixin.get_encoded_api_path() in helpers/storage_resource.py.
• Each path segment is URL-decoded then re-encoded, so a decoded path (my file.txt) and an already-encoded one (my%20file.txt, as returned by the API) converge to the same wire form without double-encoding. Absolute URLs (always encoded by the API) pass through untouched.
• Applied across:
  • files.py — upload, download, delete, move_to, copy_to, get_metadata
  • metadata.py — the shared get (covers conversations and direct low-level client.metadata.get(...) calls)
  • prompts.py — save, get, delete, get_metadata
  • resource_permissions.py — grant body urls

Verification:

• New regression tests for reserved-character encoding + already-encoded round-trip across move/copy, download (URL + decoded filename), and both files.get_metadata and low-level metadata.get.
• Full unit suite: 279 passed; pyright clean; nox -s format clean.
• Differential wire test (new vs old code): previously-working inputs are byte-identical (no double-encoding, no regression); only the previously-broken reserved-char cases changed.

Notes:

• Identifier endpoints (application/deployments/model/toolset) were left unchanged — they interpolate deployment/model names, not storage paths, and conventionally contain no reserved characters.
• Caveat of the normalize-both contract: a filename containing a literal % sequence can't be expressed from a decoded path (extremely rare).

Checklist

• Title of the pull request follows Conventional Commits specification

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[adubovik]

Let's rename it to _percent_encode_relative_url since there is nothing in the implementation of the function that's specific to DIAL resources, and I don't see why we need to make it public.

[adubovik]

Let's instead move the encoding to the very place where url is converted into the resource: safe_parse_storage_resource.
Remove get_encoded_api_path, since get_api_path will do the encoding for you.

[adubovik]

Let's extract this into a grant_body helper:

{
                    "resourcePermissions": [
                        {
                            "url": percent_encode_resource_url(url),
                            "permissions": permissions,
                        }
                        for url in resources
                    ],
                    "receiver": receiver,
                }

Reuse in sync and async versions. And use get_encoded_api_path just like we did in _move_copy_body.

[adubovik]

Let's extend the type of the argument to url: str | PurePosixPath; this will eliminate a lot of str(...) conversions downstream.