AMT coverage: g1-knownmagic-enum-prefix-b - #202

Adversarial mutation-test coverage for group g1-knownmagic-enum-prefix-b (30 behaviours across 5 units), scanned commit bba50a7.

Method: all 76 mutants were first probed against the pre-existing suite (attribution pass), then each surviving mutant got a new discriminating test and was re-probed with --only until killed. No existing test was deleted, weakened, or edited to pass under a mutation.

Behaviour matrix

Equivalent mutants (recorded, not killable)

• M53 error.offset() == len -> <=: with serde_cbor 0.11.2 an eof error's offset is always the end of input (== len), verified empirically for clean ends, truncated u64 headers, truncated arrays and missing values. The else branch is unreachable; the effective trailing-truncation guard is the final length check (now pinned by test_cbor_decode_trailing_truncated_is_corrupt).
• M56 track.len() != metas.len() -> false: both vectors are pushed in the same iteration and any push failure returns early, so the lengths can never differ at the guard.

Both are recorded in the audit issue about unreachable guard conditions.

QA

• Discriminating tests: test_address_list, test_all_discriminants_pinned, test_all_prefixes_start_with_0xff, test_try_from_u64_roundtrip_all, test_try_from_u64_unknown_magic, test_strum_kebab_case_parse_display, test_serde_kebab_case_roundtrip, test_hash_bare_vs_document, test_cbor_decode_empty_is_corrupt, test_cbor_decode_trailing_truncated_is_corrupt, test_cbor_decode_trailing_garbage_errors, test_cbor_decode_missing_payload_errors, test_cbor_decode_missing_magic_errors, test_cbor_decode_unknown_magic_errors, test_unpack_decodes_content_encoding, test_unpack_into_whitelist, test_try_into_string_invalid_utf8_errors, test_try_into_vec_decodes_encoding, test_content_encoding_deflate_roundtrip, test_content_encoding_passthrough, test_content_encoding_decode_fixtures, test_content_encoding_decode_garbage_errors, test_content_headers_strum_names, test_known_meta_try_from_magic, test_known_meta_strum_parse_display, test_normalize_op_v1_canonicalizes, test_normalize_op_v1_rejects_invalid_symbol, test_normalize_solidity_abi_v2_rejects_bad_input, test_normalize_solidity_abi_v2_canonicalizes, test_normalize_interpreter_caller_rejects_empty_methods, test_normalize_passthrough_for_binary_metas - each fails on base under its target mutation (verified by mutation-probe pass 2: every re-probed mutant reports KILLED naming these tests as killers; probe logs per mutant)
• Mutations applied: 76 exact-string mutants over crates/cli/src/meta/{magic.rs,mod.rs,normalize.rs}; the behaviour matrix above maps every mutation to its verdict and killing test. Pass 1 (pre-existing suite): 36 KILLED / 40 SURVIVED / 0 NO-RUN / 0 HARNESS-ERROR. Pass 2 (with new tests): all survivors KILLED except two mutants proven equivalent (M53 eof-offset ==-><=, M56 track/metas length check removal), documented in cbor_decode corrupt-meta guard carries unreachable sub-conditions (eof offset equality, track/metas length) — equivalent-mutant evidence #190.
• Oracle: metadata-v1 spec (magic number table, mandatory keys 0/1, 0xff prefix rule, deflate content-encoding), handwritten cbor bytes per RFC 8949 with keccak256 computed over those literals (not over cbor_encode output), out-of-band generated zlib/raw-deflate fixtures pinned as byte literals, RFC 1950 CMF byte 0x78, and the CLI's documented kebab-case names (e.g. build's default rain-meta-document-v1) — all independent of the implementation under test.
• Category check: assignment g1-knownmagic-enum-prefix-b lists 30 behaviours across 5 units (KnownMagic enum+prefix+TryFrom, RainMetaDocumentV1Item codec, ContentEncoding/Type/Language, KnownMeta+TryFrom, KnownMeta::normalize); all 30 probed and covered per the matrix — no behaviour left unprobed.

Issues filed from the adversarial pass (audit + adversarial labels): #186, #187, #188, #189, #190, #191, #203, #204, #216. (#217/#218 are accidental duplicates of #203/#204 from a re-dispatched run of this group and are marked as such.)

🤖 Generated with Claude Code

Independent re-verification (second g1 run, 2026-08-22)

A re-dispatched run of this group independently rebuilt the harness from scratch (fresh clone at bba50a7, its own 92-mutant file) and converged on the same matrix:

• Pass 1 against the pre-existing suite: 40/86 killed, 46 survived, 0 no-run, 0 harness errors — same survivor set as the original run where the mutant sets overlap.
• All survivors were then re-probed with --only against THIS branch's delivered tests (52 mutants including per-key Deserialize dispatch and per-mime rename probes the first pass did not carry): every non-equivalent mutant is KILLED by the tests on this branch; the two analytically equivalent mutants (eof-offset comparison, ContentEncoding kebab-vs-snake on single-word variants) survive as predicted and are documented in cbor_decode corrupt-meta guard carries unreachable sub-conditions (eof offset equality, track/metas length) — equivalent-mutant evidence #190.
• The eof-offset equivalence was additionally proven from serde_cbor 0.11.2 source: both SliceRead eof paths (end() and byte-exhaustion) construct the error with offset == slice.len(), so error.offset() == len is always true for eof from a slice.
• rainix-rs / static was red on formatting only; fixed by the style: rustfmt commit with no semantic change.