test_measured_from_matches_the_hdl¶

A recorded envelope must say which RTL produced it, and still match it.

Resource numbers are not self-checking. They are transcribed from a synthesis run into the registry by hand, and once written nothing in the tree can tell whether the RTL beneath them has moved since: the drift detector that would notice (synthesize_cores.parse_reports’ registered_matches) only runs when somebody runs Vivado. Two separate stale-envelope findings – row-unpack understated by ~1000 LUT at three of its coordinates, wide-field-sum-aggregation by up to 106 LUT and 141 FF at three more – trace to ONE commit with one shape: it re-measured the coordinates it ADDED and not the ones it CHANGED. Both were found by accident, months later, while attributing something else.

measured_from closes that. It stamps the recorded numbers with a digest of the HDL closure they describe (CoreDefinition.hdl_closure_digest() – the tile’s own source, the packages it imports, the headers it includes), and this module recomputes it. An edit to any of those files fails HERE, naming the tile, in the fast job, seconds after the edit – which is the property that matters: the check fires on exactly the tiles a change TOUCHED, not on the ones somebody remembers to look at.

Attributes¶

UNSTAMPED_ENVELOPES

Functions¶

`_resolve`(name)
`_measured_own_cores`() → tuple[dau_core.cores.CoreDefinition, ...]
`test_every_stamped_envelope_matches_its_hdl_closure`() → None	THE GUARD. A stamped tile's HDL must still digest to what its numbers
`test_the_unstamped_envelopes_are_exactly_the_declared_ones`() → None	Both directions, for the reason every exact set in this suite is:
`test_every_core_has_a_computable_hdl_closure`() → None	A closure that cannot be resolved is a defect whether or not the tile
`test_the_closure_holds_the_packages_a_tile_imports_and_the_headers_it_includes`() → None	The two directions that decide whether the guard works.
`test_a_generated_core_is_digested_from_its_rendered_text`() → None	A generated tile has no checked-in .sv, so a digest over files on
`test_a_moved_dependency_moves_every_dependent_tiles_digest`(...) → None	Proved by construction rather than by being green: editing a shared
`test_an_unresolvable_include_refuses_rather_than_digesting_a_partial_closure`(...) → None	A closure that quietly dropped a file it could not find would produce a
`test_an_unregistered_dependency_refuses_rather_than_digesting_a_partial_closure`() → None	The same rule one edge out: a package the registry cannot resolve is a
`test_a_stamp_without_a_measurement_is_refused`() → None	The stamp says which RTL a measurement describes; without one it says
`test_the_shell_register_block_is_stamped_and_its_stamp_matches`() → None	identity-registers is a combinational shell adapter measured on both

Module Contents¶

test_measured_from_matches_the_hdl._resolve(name: str)¶

test_measured_from_matches_the_hdl.UNSTAMPED_ENVELOPES: frozenset[str]¶

test_measured_from_matches_the_hdl._measured_own_cores() → tuple[dau_core.cores.CoreDefinition, ...]¶

test_measured_from_matches_the_hdl.test_every_stamped_envelope_matches_its_hdl_closure() → None¶

THE GUARD. A stamped tile’s HDL must still digest to what its numbers were recorded against.

Failing here does NOT mean the numbers are wrong – it means nobody has said whether they are. There are exactly two honest ways out and neither is editing the digest to match:

Re-measure the tile’s recorded coordinates through dau-build task=tasks/build/synthesize-cores and record what comes back, stamp included.
Show the edit is netlist-neutral – one control coordinate reporting envelope_drift=none against the pre-change tree – and re-stamp with that evidence in the pull request.

test_measured_from_matches_the_hdl.test_the_unstamped_envelopes_are_exactly_the_declared_ones() → None¶: Both directions, for the reason every exact set in this suite is: a stale name means a tile got stamped and nobody tightened the guard, and an unlisted name means a measurement shipped with no record of what it measured.

test_measured_from_matches_the_hdl.test_every_core_has_a_computable_hdl_closure() → None¶: A closure that cannot be resolved is a defect whether or not the tile is stamped yet – otherwise a missing header or an unregistered dependency stays invisible until the day somebody first tries to stamp that tile.

test_measured_from_matches_the_hdl.test_the_closure_holds_the_packages_a_tile_imports_and_the_headers_it_includes() → None¶

The two directions that decide whether the guard works.

Too permissive and it misses the edit that matters: a package edit must reach every tile that imports it. Too strict and every unrelated edit re-measures the world: a header a tile does not include must not reach it.

test_measured_from_matches_the_hdl.test_a_generated_core_is_digested_from_its_rendered_text() → None¶: A generated tile has no checked-in .sv, so a digest over files on disk would be blind to the one thing that decides its RTL. Its constants live in the generator and in the registry entry’s generator config; move either and the digest must move.

test_measured_from_matches_the_hdl.test_a_moved_dependency_moves_every_dependent_tiles_digest(tmp_path, monkeypatch) → None¶: Proved by construction rather than by being green: editing a shared package must fail the tiles that import it, which is the case a per-file stamp on the tile alone would miss.

test_measured_from_matches_the_hdl.test_an_unresolvable_include_refuses_rather_than_digesting_a_partial_closure(tmp_path, monkeypatch) → None¶: A closure that quietly dropped a file it could not find would produce a digest that is stable across an edit to that file – the exact failure the stamp exists to prevent, dressed as a pass.

test_measured_from_matches_the_hdl.test_an_unregistered_dependency_refuses_rather_than_digesting_a_partial_closure() → None¶: The same rule one edge out: a package the registry cannot resolve is a file synthesis would read and the digest would not.

test_measured_from_matches_the_hdl.test_a_stamp_without_a_measurement_is_refused() → None¶: The stamp says which RTL a measurement describes; without one it says nothing and reads as though the tile were checked.

test_measured_from_matches_the_hdl.test_the_shell_register_block_is_stamped_and_its_stamp_matches() → None¶: identity-registers is a combinational shell adapter measured on both parts (2026-10-06) with no clock; its stamp was written by the build host’s run and must match the RTL this checkout carries.