ocp_mx.t27: a wrong decimal in the outlier comment, and E8M0 constants copied from e8m0.t27 #7310

Two small findings in specs/numeric/ocp_mx.t27 (merged in #6828 at 322cc77). Neither changes a test result.

1. A wrong decimal in a comment (line 290)

// Inputs in units of 2^-12; entry 13 is the outlier (13517 / 4096 = 3.29999...).

13517 / 4096 is 3.300048828125, not 3.29999. The test values are right; only the comment is wrong. The worked block's text ("one outlier, 3.3") and every test assert integers, so nothing else moves. The fix: (13517 / 4096 = 3.300049, the nearest 2^-12 step to 3.3).

2. The E8M0 constants are copied, not imported

ocp_mx.t27:80-81 declares its own E8M0 constants:

const E8M0_BIAS   : i64 = 127;    // as in `e8m0.t27`
const E8M0_NAN    : i64 = 255;

specs/numeric/e8m0.t27:52-53 already declares E8M0_BIAS : i32 = 127 and E8M0_NAN_CODE : u32 = 255. Those are module-private there (no pub), with different types and a different name for the NaN code. So ocp_mx cannot import them today, and it copies them with an "as in" comment. That is the hand-copied-constant shape this codebase records as a defect: if one file changes and the other does not, nothing notices.

Options:

• (a) e8m0.t27 makes E8M0_BIAS and E8M0_NAN_CODE pub, and ocp_mx.t27 uses them. This needs the use path to work for specs/numeric (see 112 use lines name no spec: fix them before #7176 makes it an error #7191 on use lines that name no spec), plus a cast at the use site for the i32/u32 to i64 difference.
• (b) Keep the copy but make it checked. ocp_mx gets a test that pins the two values to Table 7 of OCP MX v1.0, and e8m0 gets the same test. A change to one then fails a test that names the other.

(a) is the root fix. (b) is the stopgap if use across numeric specs is not ready.

🤖 Generated with Claude Code

Boundary

• specs/numeric/e8m0.t27
• specs/numeric/ocp_mx.t27

User Scenarios

• Given specs/numeric/ocp_mx.t27 line 290 has a comment stating 13517 / 4096 = 3.29999..., when the comment is corrected, then it reads 13517 / 4096 = 3.300049, the nearest 2^-12 step to 3.3.
• Given specs/numeric/e8m0.t27 declares E8M0_BIAS : i32 = 127 and E8M0_NAN_CODE : u32 = 255 as module-private, and specs/numeric/ocp_mx.t27 copies them as E8M0_BIAS : i64 = 127 and E8M0_NAN : i64 = 255, when e8m0.t27 makes these constants pub and ocp_mx.t27 imports them with appropriate casts, then both files share a single source of truth for the E8M0 constants.

Requirements

• FR-001: The comment at specs/numeric/ocp_mx.t27:290 MUST state the correct decimal value 3.300049 (or 3.300048828125) with an explanation that it is the nearest 2^-12 step to 3.3.
• FR-002: specs/numeric/e8m0.t27 MUST export E8M0_BIAS and E8M0_NAN_CODE as pub constants so they can be imported by other specs.
• FR-003: specs/numeric/ocp_mx.t27 MUST import E8M0_BIAS and E8M0_NAN_CODE from e8m0.t27 using a use statement and apply the necessary casts from i32/u32 to i64 at the use site, removing the hand-copied constant declarations.
• FR-004: The use path mechanism MUST work for importing from specs/numeric/e8m0.t27 into specs/numeric/ocp_mx.t27.

Success Criteria

• t27c specs/numeric/ocp_mx.t27 compiles without error and exit 0
• t27c specs/numeric/e8m0.t27 compiles without error and exit 0
• The comment at specs/numeric/ocp_mx.t27:290 contains 3.300049 (or 3.300048828125) and no longer contains 3.29999
• specs/numeric/ocp_mx.t27 contains no declaration of E8M0_BIAS or E8M0_NAN constants (they are imported from e8m0.t27)
• specs/numeric/e8m0.t27 declares pub const E8M0_BIAS : i32 = 127 and pub const E8M0_NAN_CODE : u32 = 255
• specs/numeric/ocp_mx.t27 uses use specs/numeric/e8m0.t27 (or equivalent working path) and references e8m0.E8M0_BIAS and e8m0.E8M0_NAN_CODE with casts to i64