[botandbang]

What and why

leaked looked for at-<n> and rt-<n> as substrings (shared_accounts.rs#L109). an_account_reads_back_the_same_wherever_its_asked passes it everything the commands printed, which includes each sign-in link with its random state and code_challenge. Both are base64url, so about 65 random characters per link can spell at-1. By my estimate that's about 1 run in 6,000.

This may be what failed this run, where the test failed once with failed 0 of 10 replays and no panic text, but that log can't prove it: Hegel prints no panic text for a failure it can't replay. 300 cases of the property, 37 times a normal run, passed on macOS, so I found no other cause there. The run was on Linux and I didn't stress it on Linux.

Added

• holds_token in mock_oauth.rs, shared by both crates' tests: a token counts only when it isn't part of a longer word.
• a_token_is_told_from_the_same_letters_in_a_random_word. It fails with the old substring match, on state=Qat-1Zk.

Changed

• Mock::leaked in both crates uses holds_token.

Removed

Nothing.

How it was tested

On macOS: the new test, with the check broken and then restored. Both packages in one run, 3,518 tests, 3,517 passed, and the one timeout was a temporary 300-case build of the account property hitting its 120 s limit. The final holds_token differs from that run only in formatting.

[github-actions]

Rust coverage

From run 350. Both are held to the thresholds in .cargo/config.toml.

Tests	Lines	Functions	Regions
All	97.66% (87906/90016)	97.07% (14299/14731)	96.05% (194296/202294)
Hegel property tests alone	96.08% (86484/90016)	95.38% (14051/14731)	94.35% (190856/202294)