fix(agent): refuse remaining invisible metadata code points

crates/agent-identity-conformance/COVERAGE.md

@@ -14,8 +14,9 @@ The `p_*` tests check an authority and the `a_*` tests the agent; the tests name
 | §2 `u` read as text: `user@`, empty user name, `@` in path, non-ASCII host           | `a_pending_file_deleted_on_permanent_error`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |
 | §2.1 precedence of `token_url_mismatch`                                              | `p_enroll_csr_url_refusal_precedes_key_lookup_and_token_state` (over the key lookup, a revoked token and an exhausted one), `p_enroll_csr_url_refusal_precedes_token_expiry`; `p_enroll_csr_must_name_token_url` (`invalid_request` for a CSR that also has a bad self-signature)                                                                                                                                                                                                                                                                                                                                                                                                                                                                       |
 | §3 metadata limits                                                                   | `p_metadata_limits`, `p_channel_hello_invalid_metadata_rejected`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
-| §3 invisible format characters                                                       | `p_metadata_invisible_format_characters` (each end of every listed range refused in the hostname, and U+202E in an unknown key, at enrollment without consuming a token use; each code point just outside a range and the unlisted `Cf` code points U+0600, U+110BD and U+1D173 accepted and stored as sent), `p_channel_hello_invalid_metadata_rejected` (U+E007F in `Hello`); the harness test `invisible_format_character_lists_match_the_contract` pins both lists                                                                                                                                                                                                                                                                                  |
-| §3.1 invisible format characters removed, empty keys `unknown` or omitted            | `a_metadata_invisible_format_characters_removed`, through the metadata override, which also checks that the agent keeps the code points beside the list and omits an optional key (`domain`) left empty; the metadata check of every agent test refuses a listed code point in any value the agent sent, which the harness test `agent_metadata_check_refuses_listed_format_characters_only` pins                                                                                                                                                                                                                                                                                                                                                       |
+| §3 invisible format characters                                                       | `p_metadata_invisible_format_characters` (each end of every listed range, U+00AD through U+E0FFF, and U+2065, U+180D, U+FFF8, U+E0002, U+E001F and U+E0080 inside a range, refused in the hostname, and U+202E in an unknown key, at enrollment without consuming a token use; each code point just outside a range and the unlisted `Cf` code points U+0600, U+110BD and U+13430 accepted and stored as sent, in the hostname and two per unknown key), `p_channel_hello_invalid_metadata_rejected` (U+E0FFF, the last listed code point, in `Hello`); the harness test `invisible_format_character_lists_match_the_contract` pins both lists                                                                                                          |
+| §3.1 invisible format characters removed, empty keys `unknown` or omitted            | `a_metadata_invisible_format_characters_removed`, through the metadata override, which uses the range ends and the six code points inside a range as separators and as a value made only of listed code points, and also checks that the agent keeps the code points beside the list and omits an optional key (`domain`) left empty; the metadata check of every agent test refuses a listed code point in any value the agent sent, which the harness test `agent_metadata_check_refuses_listed_format_characters_only` pins                                                                                                                                                                                                                          |
+| §3.1 `machine_id` present only when the OS provides one                              | the metadata check of every agent test compares `machine_id` with the Linux `/etc/machine-id` after the sanitising of §3.1, and requires the key absent when the file is missing or leaves nothing, and compares nothing when the file cannot be read for another reason; the harness tests `machine_id_is_refused_when_the_host_provides_none`, `machine_id_must_equal_the_host_value`, `machine_id_is_not_compared_when_the_host_value_is_unavailable`, `machine_id_expectation_follows_how_the_host_file_read_ended`, `expected_machine_id_removes_control_characters_and_trims_nothing` and `expected_machine_id_is_truncated_on_a_character_boundary` pin that decision. Windows and macOS values are not compared                                                                                                                                                                      |
 | §4 certificates, §5.1 trust anchors                                                  | `p_trust_anchor_lists_roots`, `p_renew_rejects_sha384_csr`, `p_compressed_point_csr_rejected`, `p_csr_bad_self_signature_rejected`, the `csr` vectors; `p_failed_enrollment_consumes_nothing` sends the `bad_self_signature` and `ecdsa_with_sha384` vectors at enrollment, where their enrollment URL (`https://example.test/dvls`) is not the target's `u`, and §2.1 orders `invalid_request` before the `token_url_mismatch` it would draw, so it isolates their rules there                                                                                                                                                                                                                                                                         |
 | §4 chain signature algorithms, CA basicConstraints and keyCertSign                   | the chain check of `p_trust_anchor_lists_roots`, `p_reset_rebases_root_validity`, `p_renew_happy_path`, `p_renew_within_grace`, `p_rotation_publishes_both_roots_and_issues_from_new`, `p_rotation_grace_renewal_after_early_completion` and `p_rotation_old_root_cert_renewable_after_deadline`, which verifies the root's self-signature too; the harness tests `chain_verifies_p384_intermediate_and_refuses_other_algorithms` and `chain_requires_ca_extensions_on_every_issuer` pin that it accepts a P-384 issuer and refuses another curve and each missing CA extension                                                                                                                                                                         |
 | §4 leaf extensions                                                                   | the leaf profile check of `p_trust_anchor_lists_roots` (enroll), `p_renew_happy_path` and `p_renew_within_grace` (renew); the harness test `leaf_extensions_outside_the_profile_are_named` pins that the check refuses basicConstraints and names it, and refuses a leaf without the enrollment URL extension, with two, with a critical one, with one holding a `UTF8String` or naming another URL                                                                                                                                                                                                                                                                                                                                                     |
@@ -50,7 +51,8 @@ The `p_*` tests check an authority and the `a_*` tests the agent; the tests name
 - Known gaps, with no test yet:
   - the §1 conventions beyond the error body, timestamps, repeated members, nesting depth, invalid Unicode and canonical `base64`;
   - a renew or check-in body whose metadata holds a control character or an invisible format character (§3), which `p_metadata_limits` and `p_metadata_invisible_format_characters` send only at enrollment and `p_channel_hello_invalid_metadata_rejected` only in `Hello`;
-  - a code point inside a listed range of §3 other than its ends;
+  - a code point inside a listed range of §3 other than the six that `p_metadata_invisible_format_characters` and `a_metadata_invisible_format_characters_removed` send;
+  - the `machine_id` value on Windows and macOS (§3.1), which the metadata check does not compare, because the suite has no independent reading of the SMBIOS UUID or `IOPlatformUUID`, and on a Linux host whose `/etc/machine-id` the suite cannot read for a reason other than its absence;
   - the agent's removal of C0, DEL and C1 control characters (§3.1), and its removal of invisible format characters at renew, check-in and `Hello`, because no test sets such a value in the metadata override before those sends;
   - an `agent_channel_url` holding an `@`, which the agent treats as absent (§5.2, §7.2), because no mock fault delivers such a URL;
   - the shape of the agent's own CSR (§4: one attribute, an `extensionRequest` holding only the enrollment URL extension) and the URL its renewal CSR names, because the mock reports no CSR it received; the mock accepting the agent's enrollments shows only that their CSRs name the token's `u` where the mock reads it;