admin: reject non-canonical config array indices

[Amemoyoi]

This rejects non-canonical numeric array indices in /config traversal.

Previously, array path components were parsed with strconv.Atoi, which meant paths like 01 were accepted and normalized to 1. That could cause the authorization layer to accept one textual path while config traversal resolved a
different array element.

This change introduces canonical index parsing so that only normalized decimal forms are accepted:

• allowed: 0, 1, 10
• rejected: 01, 002, +1, -0

Validation:

• go test ./... -run TestUnsyncedConfigAccessRejectsNonCanonicalArrayIndices
• go test ./... -run TestUnsyncedConfigAccess

I also ran go test ./..., but the run hit an unrelated integration timeout/failure in caddytest/integration (TestACMEServerWithDefaults).

Assistance disclosure:
I used an LLM to help review the code, reason about the behavior, and help draft the original report and parts of the patch/PR text. I manually reproduced the issue, validated the fix, and ran the relevant tests myself.

[CLAassistant]

All committers have signed the CLA.

[Amemoyoi]

I checked the failing CI jobs. The failures appear unrelated to this change:

• govulncheck is reporting existing findings in ACME/tracing-related code, not in admin.go
• lint (mac) is reporting existing atomic modernize suggestions in modules/caddyhttp/reverseproxy/*

The targeted tests for this change are passing locally:

• go test ./... -run TestUnsyncedConfigAccessRejectsNonCanonicalArrayIndices
• go test ./... -run TestUnsyncedConfigAccess

[mholt]

Thanks. Can you please sign the CLA and restore the deleted assistance disclosure from the PR template?

[Amemoyoi]

I’ve signed the CLA and restored the assistance disclosure in the PR description.

It shows “You have agreed to the CLA for caddyserver/caddy” on my side, so it looks like the signature is in place, but the CLA check may not have updated yet.

[mholt]

The CLA needs to be signed by the same email address as on the commits.

[Amemoyoi]

I checked it and found that the commit email was the issue. I’ve updated it to match the CLA-signed address and pushed the updated commit to the PR branch. Thanks.

[francislavoie]

LGTM, could use a couple more test cases to cover more of the allowed and rejected cases though (including +/- as described)

[Amemoyoi]

Thanks. I added broader test coverage for both allowed and rejected forms, including the +/- cases, and pushed the update.