Harden URL/DHT parsing, fix repo write credential, load nextest config

[tripledoublev]

Summary

Four safe, self-contained fixes from a review of the group/repo key handling. No public API signatures change.

• Parsing panics → errors. public_key_from_query / secret_key_from_query / shared_secret_from_query used copy_from_slice into a [u8; 32], panicking on a wrong-length value from an attacker-controlled join URL. Replaced with a shared key_bytes_from_query helper that errors on bad hex or wrong length.
• Decrypt panic → error. decrypt_aead sliced data[..24] with no length check, panicking on a short/hostile DHT value. Added a guard that returns an error when the value can't carry a nonce.
• Repo write credential. create_repo built the in-memory Repo with the group's owner secret instead of the repo record's own. Now uses repo_dht_record.owner_secret(), matching the CommonKeypair persisted alongside it.
• Nextest config actually loads. The root nextest.toml was never read (nextest loads .config/nextest.toml) and used the invalid test-timeout key. Moved to .config/nextest.toml and switched to slow-timeout = { period = "600s", terminate-after = 1 }. It now loads with no warnings, enforcing the intended 10-minute hard per-test timeout.

Notes

• --test-threads=1 --retries=3 in CI already matched the config values; the new effect is the hard per-test timeout now taking effect.
• Architectural items from the review (owner-secret in invite URL, backend mutex held across network I/O, slot-count repo advertisement races) were intentionally left out of this PR.

Verification

• cargo build, cargo fmt --check, cargo clippy --all-targets --all-features -- -D warnings all pass.
• cargo nextest list auto-loads .config/nextest.toml with no warnings.

🤖 Generated with Claude Code