Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions changelog.d/6606-pex-runset-net-correspondence.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
- **`klt pex --engine klayout-runset --deck <curated>` joins runset devices with curated parasitics by geometric net correspondence** (#6606, 2026-11-20). The runset leg is re-extracted as usual, a curated deck runs with parasitics, and the two are joined on strictly interior metal-geometry probe points (layer-identity matched, ≥99% by area); the curated lumped R/C network is merged into the leg. Unmatched nets and split/merge disagreements between the two engines are disclosed in `extraction.net_correspondence`, never silently dropped or merged. The measured decision and the refuted alternatives: `docs/design/runset-pex-net-correspondence.md`.
- **`klt signoff` refuses a parasitic-free pex leg for item 7** (#6606, 2026-11-20). A citation whose envelope discloses `extraction.parasitics_carried: false` (a plain `--engine klayout-runset` run) renders `unmet` with the new reason `no_parasitics` — the remedy is re-running the same artifact with `--deck`; a merged leg grades exactly as a curated one. The `pex_extraction` disclosure itself is unchanged.
21 changes: 12 additions & 9 deletions docs/cli/extract.md
Original file line number Diff line number Diff line change
Expand Up @@ -699,16 +699,19 @@ rather than silently ignored.

**No parasitics.** The result is a device-level netlist with
`parasitics: null`. Combining runset devices with curated-deck parasitics
would need a geometric net correspondence between two independent extractions
needs a geometric net correspondence between two independent extractions
that number their unlabelled nets independently (`\$1` above; see "Anonymous
net numbering" below) — nothing in this repo provides that today
([#6606](https://github.com/2AMLogic/klayout-tools/issues/6606)).
[`klt pex --engine klayout-runset`](pex.md) (issue #6559) does use this engine
for its extracted leg, as a **device-level, parasitic-free** leg whose envelope
says so (`extraction.parasitics_carried: false`, `extraction.params_not_carried`).
That route does **not** by itself make `docs/design-evidence-tiers.md` T1 item
7 citable for such a block: item 7 is parasitic verification, and a runset leg
carries none. #6606 is what closes that.
net numbering" below) — this engine stays device-only, and `klt extract
--engine klayout-runset --deck` is refused as ever.
[`klt pex --engine klayout-runset --deck <curated>`](pex.md) (issues #6559,
#6606) is where that correspondence lives: it runs this engine for devices,
a curated deck for parasitics, joins the two by measured geometric rules
([`../design/runset-pex-net-correspondence.md`](../design/runset-pex-net-correspondence.md)),
and merges the parasitic network into the leg it re-simulates. A `klt pex`
report from that route can satisfy `docs/design-evidence-tiers.md` T1 item
7's intent when its envelope says `extraction.parasitics_carried: true`;
a parasitic-free leg (either runset shape) is refused by `klt signoff`
(`no_parasitics`, issue #6606).

**Report.** Additive fields only, so `SCHEMA_VERSION` is unchanged (4): the
report carries a top-level `engine`, `deck_file`, `deck_vars`, `deck: null`,
Expand Down
114 changes: 86 additions & 28 deletions docs/cli/pex.md

Large diffs are not rendered by default.

25 changes: 24 additions & 1 deletion docs/cli/signoff.md
Original file line number Diff line number Diff line change
Expand Up @@ -4061,6 +4061,29 @@ extraction-only report can never render item 7 `"met"`; see "A check that
checked nothing is refused, not reported" above for the shared rule and both
routes into it.

**`no_parasitics` vs. all of them** (issue #6606, AC 4). A fourth thing can
be wrong with an item-7 citation: the cited `klt pex` report's own
extraction disclosure states the extracted leg carried **no parasitics**
(`extraction.parasitics_carried: false` — a plain
`--engine klayout-runset` leg, issue #6559). Item 7 is parasitic
verification, and such a leg's `delta[]` rows are device-representation
deltas, so the verdict does not establish the item no matter that every row
met tolerance: it renders `"unmet"` with `reason: "no_parasitics"`, whose
remedy is "re-run *this* artifact with parasitics attached"
(`klt pex --engine klayout-runset --deck <curated>`, which joins the curated
deck's parasitics onto the runset devices by geometric net correspondence —
[`../design/runset-pex-net-correspondence.md`](../design/runset-pex-net-correspondence.md)).
A merged leg that attached cards discloses
`extraction.parasitics_carried: true` and grades exactly as a curated one;
an envelope with no engine statement (every curated-deck run, and every
record predating the route) carries parasitics by construction and is never
refused here. The gate applies only to item 7. The check keeps its own
`passed` status in aggregation mode. A merged citation also quotes the
ground capacitors’ `substrate_return` when present. A null `runset` identifies
a separate DC-tied return whose connection to a runset device body is
unverified. Item 7 grades the simulated leg with that disclosed return.
Item 2 (Layout) keeps its existing treatment of either leg.

Only the literal JSON boolean `true` in `environment.sdf.annotated` grants
annotation credit. False, null, missing fields, strings (including `"true"`
and `"false"`), numbers, arrays, and objects do not qualify. Malformed optional
Expand Down Expand Up @@ -5675,7 +5698,7 @@ resolved hash renders `unmet`/`unverifiable_provenance`.
| `runset_rehash_status` | string | `"unverifiable"` appears on a pinned PDK-runset LVS citation when signoff cannot hash the report's top-level `deck_file` (issue #6452). Look for it in item 4's `citation` or in the LVS entry in item 11's `citation.parts`. Signoff resolves relative deck paths from its own working directory. A missing, non-string, or unreadable path allows an otherwise valid item to remain `met` with this disclosure. |
| `content_hash_unresolved` | object | Present when a `yield` report names a samples document and signoff cannot resolve its hash (issues #2197 and #5963). `samples` is the report's reference and `searched` lists the disk paths tried, in order. If a declared store manifest also fails to supply the digest, `artifact_manifest` names its path and `reason` is one of `manifest_unreadable`, `manifest_invalid`, `member_missing`, `duplicate_member`, or `ambiguous_member`. A passing unpinned entry carries the object on its citation. For a pinned unmet item it appears in `detail.content_hash_unresolved`. The field is absent when a digest resolves or the report names no samples document. |
| `exit_status` | integer | `0`, *inferred*, for a file-backed entry (a readable, passing envelope implies its producing command exited zero); the subprocess's *actually observed* return code, for a command-backed entry. |
| `body_bias` | object | **`pex` citations only**, and only when the cited envelope carries a `body_bias` block (issue #1983): `{"status", "unbiased_device_count", "unbiased_nets"}`, reduced from it — whether the extracted netlist these post-layout numbers were measured on had a DC bias path for every device body. `status: "not_assessed"` (issue #6559: a `klt pex --engine klayout-runset` leg, which never runs the check) carries `null` for both `unbiased_device_count` and `unbiased_nets` — not `0` / `[]` — and the text rendering prints no count for it; the same citation then also carries `pex_extraction` (`{"engine", "parasitics_carried", "params_not_carried"}`, quoted from the cited envelope's `extraction`; absent for a curated-deck run), rendered as a `pex extraction:` text line beside `body bias:`. **Absent** for any other kind, and for `pex` evidence committed before `klt pex` reported it — an absent `body_bias` means "this artifact made no body-bias statement", never "every device body was biased". See "Device-body bias is reported, not graded" above. |
| `body_bias` | object | **`pex` citations only**, and only when the cited envelope carries a `body_bias` block (issue #1983): `{"status", "unbiased_device_count", "unbiased_nets"}`, reduced from it — whether the extracted netlist these post-layout numbers were measured on had a DC bias path for every device body. `status: "not_assessed"` (issue #6559: a `klt pex --engine klayout-runset` leg, which never runs the check) carries `null` for both `unbiased_device_count` and `unbiased_nets` — not `0` / `[]` — and the text rendering prints no count for it; the same citation then also carries `pex_extraction` (`{"engine", "parasitics_carried", "params_not_carried"}` plus `substrate_return` when present, quoted from the cited envelope's `extraction`; absent for a curated-deck run), rendered as a `pex extraction:` text line beside `body bias:`. **Absent** for any other kind, and for `pex` evidence committed before `klt pex` reported it — an absent `body_bias` means "this artifact made no body-bias statement", never "every device body was biased". See "Device-body bias is reported, not graded" above. |
| `parts` | array\<object\> | **Item 11 citations only** (issue #2025): every artifact of the compound cited set, each in this same citation shape (minus `parts`/`power_delivery`), in `erc`/`lvs`/`place-and-route` order. The top-level fields above describe the *leading* (`erc`) part, so a consumer written before item 11 existed still reads a well-formed citation; nothing a reader needs is reachable only through `parts`. For a **grouped** citation (issue #5597) these are the *leading group's* parts, and every group's own parts live in the item-level `citations` array (one compound citation per group, in manifest order) beside `citations_graded` and `members_covered`/`members_total` — see "Item 11 is compound" above. **Absent** for every other item. |
| `power_delivery`| object | **Item 11 citations only** (issue #2025): `{"partition_kind", "supply_nets", "pdn", "strap_layers", "tapcell_master", "power_connectivity_status", "ties_checked_by_assertion", "ties_checked_by_well_assertion", "supply_unlabelled_islands"}`, plus `tapless: true` (issue #4199) only when the cited P&R response states a tapless library (`tapcell_master` is then null by design), plus `disclosed_undeclared_tie_classes` when present, plus `power_ports` (issue #2839) only when the provisional port-accessibility half was graded (see "Item 11 port accessibility") — what the grading actually resolved, so a `met` verdict states which supplies were declared and which branch proved them. `pdn` is `false` (with `strap_layers: []`, `tapcell_master: null`) for an analog or full-custom block that cited no `place-and-route` response — "no PDN citation", not "a PDN was checked and found missing", which renders `unmet`/`no_pdn` instead. `ties_checked_by_assertion` (array\<string\>, issue #2234) quotes the cited ERC run's own `erc_coverage.checked_by_assertion`: the `erc.missing_tie` work identities whose tap region came from a caller assertion (`ties[].tap_boxes`) rather than PDK-marker narrowing — `[]` for a purely marker-derived run, and for ERC evidence produced before that field existed. It does not change the verdict (an asserted tie is graded `met` exactly as a marker-derived one, and a degenerate or unmatched assertion is rejected by `klt erc` itself); it states which taps rested on the caller's word, without re-opening the cited envelope. `ties_checked_by_well_assertion` (array\<string\>, issues #2255 and #2540) quotes that run's `erc_coverage.checked_by_well_assertion` the same way, for the **well** side: the work identities whose well side rested on caller-named coordinates, in either of two forms — the substrate region itself asserted (`ties[].well_layer: null` + `ties[].well_boxes`) because the block draws no well/tub layer at all (the native-substrate case), or a *drawn* well whose class selection was named in boxes (`ties[].well_requires_boxes` / `ties[].well_excludes_boxes`) because no drawn layer separates the tub's two bias classes. `[]` for a run using neither, and for ERC evidence produced before those fields existed. Kept as a separate list because it is a separate, weaker claim: one says which drawn geometry is the tap, the other says how the well side was named. It does not change the verdict either — `klt erc` rejects an assertion indistinguishable from the whole top-cell extent as `degenerate_well_assertion`, and a box selection that kept every shape of the drawn layer or none of them as `degenerate_well_selection`, both landing in `erc_coverage.skipped` where this item already refuses to read them as a clean missing-tie verdict. A purely marker-layer selection (`well_requires`/`well_excludes`) stays out of this list: the stream draws its own partition. `supply_unlabelled_islands` (object, issue #2524) is the **severed-rail negative** this item could not previously state: `{<declared supply name>: <nets[].unlabelled_islands>}` for each cited supply whose ERC spec declared the `stackup` roles it owns (`nets[].roles`, issue #2510), every value `0` for a met item. `erc.unconnected_net` counts islands *carrying the declared label*, so a single-label rail severed into a labelled piece and an unlabelled orphan grades clean — this key is what says the clean read was measured over the whole conductor on the owned roles, not only over what carried a label (its non-zero counterpart is `erc.unlabelled_conductor`, already an `unmet`/`supply_not_continuous` blocker in the table above). `{}` when no cited supply declared `roles`, and for every ERC envelope produced before those keys existed — the honest "not measured", never a fabricated zero, so a grader can tell "checked, and it is zero" from "nobody asked". Like the two assertion lists, it does not change the verdict; it states the negative the verdict rests on. `disclosed_undeclared_tie_classes` (array\<object\>, issue #2623) is the **partial-declaration** counterpart of `detail.ties_disclosure_reason` below: `[{"class", "reason"}, …]`, one entry per `erc.missing_tie:["<class>"]` the cited ERC run recorded in `erc_coverage.inapplicable` with a disclosed reason (`ties_disclosure.undeclared_classes`, issue #2541) — a well class the spec names as inexpressible or tool-limited *alongside* at least one other tie it did declare and get checked (the `tie_count > 0` path this item is already `met` on). It does not change the verdict — the declared, checked tie(s) already carry the item to `met` on their own merits, and a disclosure is never a substitute for a computed result — it only states, in the report of record, that a second class was considered and why it could not be checked. **Absent** (not an empty list) whenever the cited run disclosed no named class, so a run predating #2541 or naming none renders this citation exactly as it did before this key existed. These are **per-stream facts** (issue #5597): on a grouped citation every group's `citations` entry carries its own `power_delivery` block, and the values are never merged across groups — each stream proves its own grid. |
| `coverage` | object | **`drc` citations only**, and only when the cited envelope carries a `coverage` block (issue #2002): `{"layers_in_stream_without_rules", "rules_skipped", "deck_scope"}`, quoted verbatim from it — the three fields [`design-evidence-tiers.md`](../design-evidence-tiers.md) item 3 requires a DRC claim to disclose. **Absent** for any other kind, and for DRC evidence committed before `klt drc` reported coverage — an absent `coverage` means "this artifact reported no coverage", never "this deck has no gaps". See "DRC coverage is reported, not graded" above. |
Expand Down
11 changes: 10 additions & 1 deletion docs/design-evidence-tiers.md
Original file line number Diff line number Diff line change
Expand Up @@ -480,7 +480,16 @@ written to every block.
proof" section) — every `delta[]` row is explainable directly from the
extracted netlist's own added R/C, not merely reported, including the
(measured, not assumed) no-degradation rows that canary's current
topology produces.
topology produces. For a block whose core device family only the PDK's
own LVS runset recognises (the case `klt pex --engine
klayout-runset` exists for, #6559), the citation must carry parasitics:
`klt signoff` refuses a `parasitics_carried: false` leg with
`no_parasitics` (#6606) — the merged route (`klt pex --engine
klayout-runset --deck <curated>`, which attaches the curated deck's
parasitics to the runset devices through the measured geometric net
correspondence of
`docs/design/runset-pex-net-correspondence.md`) is the runset-engine
shape that grades.
- *Digital* — the functional test suite re-run against the post-route
gate-level netlist with back-annotated SDF timing, not only the
pre-layout RTL/gate simulation. The machine-checkable evidence is a
Expand Down
Loading
Loading