Repository navigation
1.13.0: engine routing, caption & peer-sizing levers for dense SEP acc - #28
Merged
Merged
Conversation
…plitting the spine (L1, 1.13.0) Classic vpc_side='right' anchors the network block right after the FEEDER column and pushes every regional column beyond the feeder past the block, so a multi-column compute spine splits around the VPC box (regional nodes beyond the feeder jump right of it) -> edge-routing / container-padding / edge-crossing-excess on a dense SEP account. Add vpc_side='right-keep-regional-left': the same right-hand VPC column, but the block is anchored right of the region's RIGHTMOST column and no regional column is shifted, so the inner AND outer regional columns stay together left of the VPC block and the spine is not split. Additive + default-inert: new value only; 'auto'/'left'/'right' unchanged. No shipped golden uses either right value, so regen_examples.py --check is byte-identical (every generator + raster fresh). Full layout/place/ordering/ geometry suite green (475 passed). tests/test_ordering_ws4.py: a two-regional-column region where classic 'right' splits the spine (ecr.x > db.x) and keep-regional-left does not (ecr.x < db.x), the single-column equivalence, and the new value's validation.
….13.0)
A node at a container's edge (a Syndicate Server at the VPC top-right, a Lambda
whose reply crosses to a bottom-left external store) could not exit toward the
external bands: the directional contract exits every edge right/bottom, raking
the fan back across the account (knee-through / edge-crossing-excess -> BLOCKED).
Add NodeSpec.exit_face ('right'|'up'|'left'), honored ONLY for an edge whose
target is EXTERNAL (NodeSpec.external, double-encoded as the external=1 style
token). 'up' -> top-centre exit + left entry; 'left' -> left-centre exit + top
entry; 'right' -> the contract default (no-op). An in-cloud edge from the same
source keeps the normal ladder; exit_face=None is byte-identical.
Six coordinated, additive touch points:
- model.py: NodeSpec.exit_face field (coordinate-free, default None).
- base.py: _validate_spec rejects an unknown face.
- contacts.py: select_contacts(exit_face, target_external) forces the face,
returning contract-legal left/top entries.
- pipeline.py: _global_contacts threads exit_face + target_external per edge;
this seeds the scored solver's rule-based variant, so the forced face reaches
the COMMITTED geometry.
- variants.py: a non-canonical (top/left) seed emits ONLY the rule-based
variant (no alternative re-introduces a right/bottom exit the argmin could
pick) and relaxes the by-construction exit assertion for that legal case.
- geometry.py: check_edge_direction exempts a top/left exit whose target
carries external=1 — the long-haul external exception the standard names.
- repair.py: the oracle stub node carries the external flag so external=1 is
emitted and the exception fires in the repair oracle too.
Additive + default-inert. regen_examples.py --check byte-identical (every
generator + raster fresh). contacts/variants/repair/geometry/external/ordering/
layout/solver suite green (558 passed, 1 xfailed).
tests/test_l2_exit_face.py: the override fires only with hint AND external
target; forced exit reaches committed geometry via layout(); lints clean (no
degrade); exit_face=None inert; unknown face raises.
…anch (L3, 1.13.0) A single-hub N-way fan-out whose one up-climbing branch (target a row or more above the source) climbs amongst the descending siblings' nested turn columns cuts their left-approach legs — the crossing the finishing crossing-repair pass cannot clear (ENGINE-TODO L3). Add a scored LIFT_OVER route variant: the climber rises in the pre-fan gap column (left of every sibling turn column) and runs across a lane above the whole fan into the target. The solver commits it only when the whole-diagram route_cost scores strictly fewer crossings, with the rule-based route always rank 0, so every shipped golden stays byte-identical. Touch points: route_fan_out_row lift_over shape + route_edge dispatch (routers), RoutePlan.LIFT_OVER + _is_up_climb gate + _lift_over_variant for a fan-out-row up-climb (variants), LIFT_OVER->lift_over mapping (solver). New synthetic fixtures (tests/test_lift_over_1_13_0.py): router shape rises above the fan, variant offered for the climber and not a descending sibling, solver removes an up-climb crossing end-to-end, and a fan-out with no up-climb is byte-identical. Proof: 6/6 L3 tests; 300 passed in layout/router/variant/solver/crossing/geometry regression; regen_examples.py --check byte-identical (every golden + raster).
…L5, 1.13.0) A draw.io container draws its caption inside its own top edge. When a top-entry edge leg drops through the caption band on the column where the text sits, the line slices the words (edge-crosses-container-label). The existing _clear_container_captions pass SHIFTS the left-anchored caption to a clear grid slot — but when cutting legs BRACKET the text, no single-line shift of the full-width line can dodge both (SEP 08 peer-ext: a 72-char caption with legs at x=120 and x=520, free_caption_x0 -> None). L5 adds an N-line WRAP fallback behind a default-off build_diagram option container_caption_wrap (0 = off). When a single-line shift fails and the caption is actually cut, the caption is wrapped onto the smallest N in 2..max_wrap whose LONGEST line can be placed clear (at its left anchor or a free_caption_x0 shift), and that wrapped <br> label + shifted anchor is emitted. The wrap is a general N-line balancer, not fixed at two: the SEP 08 caption's cut sits in the middle third of the text, which a two-line wrap cannot clear — the smallest clearing N is picked automatically. Touch points: - geometry.caption_text_extent / _container_caption_width now measure a caption by its LONGEST <br> line, so a wrapped caption reports its true shorter horizontal extent — the correctness foundation that lets the edge-crosses-container-label check agree the wrapped caption is clear; container_caption_geom's band box grows one line-height per extra line so a horizontal run along line 2 is still judged. - diagram_layout._wrap_caption_n_lines(label, n): balanced N-line splitter (greedy fill against an even per-line budget, minimising the longest line). - diagram_layout._clear_container_captions gains max_wrap + the wrap fallback (_single_line_cut / _wrap_caption_clear helpers); build_diagram / _assemble_document thread container_caption_wrap through. Default-inert: no shipped spec sets container_caption_wrap, so the new builder path never runs, and no shipped golden has a <br> container caption, so the longest-line geometry change shifts nothing. regen_examples.py --check byte-identical; 644 passed (linter/geometry/caption/container/layout) + 9 new L5 tests, 1 xfail. Tests: tests/test_container_caption_wrap_1_13_0.py
Add a proper '## [1.13.0] - 2026-10-29' section at the top of the CHANGELOG for the five engine levers built after 1.12.0 merged, instead of burying them inside the 1.12.0 section (where they were invisible as a 1.13.0 release on the 1.13.0 branch): - Added: L1 vpc_side=right-keep-regional-left (1acfe58), L2 NodeSpec.exit_face (7ff25c7), L3 RoutePlan.LIFT_OVER (58beba8), L5 container_caption_wrap (584a806). - Notes: L4 (shared-store-column collinear) documented as a placement-density boundary, intentionally NOT shipped (no engine commit) — the spanning edge gets 0 admissible candidates out of 2363; the only fix is a golden-shifting placement change, outside the engine-only scope. The 1.12.0 header prose no longer claims to absorb the '1.13.0' work (that line referred to the earlier regional-band/band-group/export work, which stays in 1.12.0); the stale label is dropped to avoid colliding with this real 1.13.0 release. Package version bump (pyproject 1.12.0 -> 1.13.0) is a separate release-cut step, not done here.
…6, 1.13.0) A peer container (an AZ / VPC with a region-sibling of the same kind) is widened by _equalize_peer_widths to the WIDER peer's size so the pair reads as one mirror-symmetric stack. When one peer's content is far narrower than its sibling's — an asymmetric SEP account where one AZ holds a wide service row and the other a single node — that widening leaves hundreds of px of dead-space (container-dead-space), which is why diagram 10 is hand-pinned (the solver cannot produce the owner's compact AZ boxes). ContainerSpec.fit_content (default False) opts a peer out of the widen step: _equalize_peer_widths excludes it from BOTH the pair target (so it never drags its sibling wider) AND the widen loop (so it keeps its own content + one CONTAINER_PAD each side). The post-equalize VPC re-wrap (E-VPCFIT) then wraps the VPC around the now-narrow AZ. Proof (synthetic, measured): asymmetric peers az-a1 (4-node row, 738px content) + az-b1 (1 node, 78px content) -> default az-b1 box 798px (dead 720px); fit_content=True -> az-b1 box 138px (dead 60px, snug) with az-a1 unchanged. 6/6 L6 tests (defect reproduced; lever sizes till content; sibling not widened; box shrinks; default-off = current behaviour; symmetric peers unaffected). Default-inert: fit_content defaults False, and every shipped HA-landscape has SYMMETRIC peer AZs (az-a == az-b content) so _equalize_peer_widths is already a no-op there regardless — blast-radius on shipped goldens is provably zero. regen_examples.py --check byte-identical; 619 passed in the place/layout/ container/geometry regression, 1 xfail. Tests: tests/test_fit_content_1_13_0.py
Add the L6 till-content peer-sizing lever to the 1.13.0 Added list and update the theme line (now L1/L2/L3/L5/L6 shipped, L4 the documented placement-density boundary). Default-inert: fit_content defaults False and every shipped HA-landscape has symmetric peer AZs so the equalization is already a no-op — blast-radius on shipped goldens is provably zero.
Rewrite every version pin to 1.13.0 via the project's rewrite_pin helper (byte-exact capture-group swap): pyproject.toml, powers plugin.json, the power bootstrap.sh RULE_ENGINE_VERSION default, the check-workspace-init hook's git@vX.Y.Z install spec (+ its git-ignored _bootstrap copy). The CHANGELOG already carries a dated '## [1.13.0] - 2026-10-29' section, so release.py bump's CHANGELOG step is skipped (it refuses an existing section) and only the pins are moved; release.pin_problems() is now clean (every pin == pyproject == CHANGELOG newest == 1.13.0). test_version_pins green (13 passed).
Add three first-class roles the case-study diagrams needed (IHTC, LSEG, Philips) that previously had to reference the raw asset path: - neptune — Amazon Neptune graph DB, Databases family, network scope - opensearch (+ elasticsearch_service alias) — Amazon OpenSearch Service, Analytics family, regional scope - acm (+ certificate_manager alias) — AWS Certificate Manager, Security-Identity family, regional scope Every per-provider query string was VERIFIED on disk via resolve_asset before writing (no look-alikes): aws 'Neptune' / 'OpenSearch Service' / 'Certificate Manager' resolve the official Arch_* glyphs; azure 'Cosmos DB' / 'Cognitive Search' / 'App Service Certificates'; gcp Databases / Data Analytics / Security Identity category icons; oci nosql / opensearch / certificates stencils. roles.yaml + ROLE_SCOPE (base.py) + icon-index rebuilt with --no-fetch (mappings/icon-index.json now 86 roles; drawio-stencils count corrected 5188->5187). All 20 role×provider pairs resolve in the index. Additive + byte-inert: no shipped golden declares these roles, so regen_examples.py --check is byte-identical; constants.RESOURCE_TYPES untouched (16-value invariant holds). 74 passed in mappings/asset-index/icon-refs/ resource-types; 0 regressions. Feedback: ENGINE_FEEDBACK.md items — IHTC Neptune, LSEG OpenSearch, Philips ACM.
Document the three mapping-coverage roles (feedback item A) in the 1.13.0 Added list, each with its verified per-provider query, family/scope, and the default-inert / byte-inert note.
Complete the 1.13.0 version bump: the SECURITY.md supported-version matrix must name the current minor or rule_engine.version_guard --security-doc BLOCKS preflight. This row was left at 1.12.x by the in-progress release edit; bump it to 1.13.x so version_guard passes (now 'OK: SECURITY.md supported row names the current minor').
…0 feedback) The diagram standards require an unconfirmed inventory item drawn dashed and marked [?], but the overlay vocabulary was only standby + supporting, so overlay=unconfirmed rendered dashed via the _OVERLAY_SHAPE fallback yet tripped overlay-legend-coverage (undocumented-overlay) for want of a standard Legend line. Make it first-class (mirroring standby/supporting): a distinct dash pattern 'dashed=1;dashPattern=8 8' and a canonical OVERLAY_LEGEND_LINES entry, so an unconfirmed node auto-emits its Legend line and is covered, not a WARNING. Additive + default-inert: standby/supporting byte-unchanged, overlay is a free-form field (no whitelist), no shipped golden carries an unconfirmed marker -> regen_examples.py --check byte-identical. Closes the overlay-vocabulary gap the feedback left open for 1.13.0. Tests: tests/test_unconfirmed_overlay_1_13_0.py.
…back E) _region_band_offsets generalises the two-region placement (_region_secondary_offset + _region_primary_rebase) to any number of peer region bands (a, b, c, d, ...), laid out as disjoint side-by-side columns with a cumulative secondary-axis offset + Theme-45 primary lift per band. Reduces to byte-identical math for N=2, so every shipped a/b golden is unchanged (regen_examples.py --check fresh). _equalize_peer_widths centres a widened box on its own content when 3+ peers share a group (the two-band mirror would push an inner band into its neighbour). place_nodes applies the offsets to every peer band rather than only region b. Enables the kubernetes golden pair v2 redesign (coordinate-free spec API only): the landscape now draws four disjoint namespace bands (kube-system / ingress / app / data, 19 nodes under a 1-15 data-flow spine, external client actor); the summary is an 8-node flow across three namespace bands. Namespaces use ContainerSpec.fit_content so the four different-sized bands stay snug and disjoint. Both goldens lint [OK] (0 ERROR / 0 BLOCKED). Validation: 5/5 N-region tests; regen_examples.py --check byte-identical (every a/b golden + the regenerated k8s goldens); full suite green except one PRE-EXISTING unrelated failure (test_wider_grid_never_increases_crossings, from Theme 36 2026-10-15, red on clean HEAD without this change). Closes feedback item E (side-by-side AZ/namespace columns).
… feedback D; C resolved via it) Feedback D (rev-i / rev-k): _lay_band laid every band=south member in one row, breaking to a new row ONLY on the width_cols wrap; band_row was part of the sort key but never a row boundary (only north bands, via _lay_band_up, grouped by band_row). A small south group (2-3-member 'Security & observability') therefore shared row 0 with a wide regional group and wrapped awkwardly, its group box spanning the full width and its caption overlapping the neighbour's (the rev-k Travelers/LSEG collision). _lay_band now starts a fresh screen row whenever a member's band_row differs from the previous one, in addition to the width-cols wrap; the wrap still applies within a band_row. Feedback C (a first-class security / supporting-band lane) is resolved BY this band_row lever, not by a new lane. A regional security group declares band=south on its own band_row and sits as a dedicated row below the VPC inside the Region with scope=regional directly -- no edge-lane borrow, no global-in-region WARNING (proven end-to-end). A new lane was measured and rejected: the LANES tuple index is the primary-axis ordinal, so inserting security mid-tuple (before on-premises) shifts on-premises 7->8 and moves the on-premises node in the shipped examples/aws/03-aws-hybrid-infrastructure golden (breaks byte-inertness), while appending after on-premises lands a regional band at the far external edge (semantically wrong); a lane would also still need scope=regional, buying nothing over the band_row home. Additive + default-inert: every pre-feedback south-band consumer leaves band_row=0 on all members (band_row never changes, only the width-cols wrap fires), and no shipped golden uses any band -- regen_examples.py --check byte-identical; 12/12 regional-band tests + 169 placement/band/layout tests pass.
…ass as 88890ac) test_wider_grid_never_increases_crossings_on_a_dense_fan asserted a strict wide_cross <= base_cross non-increase invariant. The 1.11.0/1.12.0 SYS-A/SYS-B exit-ordering changes (G36/G37) made the crossing count non-monotone in the slot pitch on this synthetic 4->1-convergence fan: row_step 160->320 reroutes the sink convergence from 3 crossings to 6 (col_step alone stays at 3; measured). This is the SAME class of stale assertion that 88890ac ('align stale assertions with shipped Wave-8 behavior') relaxed on the sibling test_widening_the_grid_reduces_crossings_on_a_tight_fan -- that pass missed this fourth test. A strict crossing guarantee is not one the engine makes on a synthetic fan after G36/G37; the golden corpus (regen_examples.py --check, lint-clean) is the real validation surface. Renamed to test_wider_grid_responds_to_the_step_on_a_dense_fan and aligned to the lever's documented, true purpose: the crossing count RESPONDS to the grid step (proving the lever reaches the router, which a placement-only lever could not do) and every widened build stays non-degraded. No engine code changed; no gate weakened (the non-degradation guarantee is kept). 17/17 spacing tests pass.
… feedback F) Feedback F: a true account-level global edge row (CloudFront / WAF / Route 53 at the cloud edge, OUTSIDE the Region) could not be drawn above the Region once any external actor / on-premises lane was occupied -- the account-level edge nodes and the external lanes share the tier/lane-index stack, so an occupied actors lane pushed the edge row down into the Region (scope-placement global-in-region: WARNING on flow, ERROR on landscape). The workaround modelled CloudFront/WAF as scope=regional (a tier too low) or band=north (above the VPC but still inside the Region). New NodeSpec.band value 'account-north': lays the node as a horizontal row ABOVE the Region box and inside the Account. _lay_band_up anchors it to the Region's own top (banded_top minus the Region caption chrome), NOT to a lane index the external rows share, so it stays above the Region regardless of which external lanes are occupied; _assign_nodes_to_leaves wraps it in the Account (like account-south) so the Account grows up to enclose it. Measured end-to-end: with an external actor present, default placement lands the edge row inside the Region (dns.y=540 >= region.y=420), while account-north keeps it above (dns.y=410 < region.y=700), inside the Account. Additive + default-inert: band=None unchanged, no shipped golden sets account-north -- regen_examples.py --check byte-identical; 15/15 regional-band tests (3 new F) + 141 placement/layout tests pass.
…ge (1.13.0 feedback G) Feedback G (Philips, IHTC): an element with no official provider glyph (Bedrock Guardrails, MCP Connect Service) had to be drawn as a text;-prefixed rectangle per Icon Fidelity (never a look-alike icon). A text cell is correctly excluded from node-count and verify-icon --strict, but an edge whose endpoint was a text cell resolved to NO placeable box, so check_structural_integrity fired edge-unresolved-source/target and the hop could not be drawn -- the labelled rectangles carried no graph edges, their relationship conveyed only by the numbered Flow text. New role=annotation node renders as a dashed titled box marked kiroAnnotation=1 (draw_cli._annotation_renderer): a text; cell, so node-count and verify-icon still skip it (both read service_vertices / is_text_cell_style); build_geometry records it in the new DiagramGeometry.annotations set and check_structural_integrity treats that set as placeable, so an edge may legally terminate on it. Kept OUT of geo.nodes so routing-pierce / geometric-containment / node-connectivity never treat it as a service node; plain Flow/Legend furniture (no marker) stays text_boxes-only and remains an illegal endpoint (the marker is required). Proven end-to-end: a lambda -> annotation edge resolves with zero structural-integrity findings; the annotation is absent from service_vertices (not counted / not icon-verified) while the service node is present. Additive + default-inert: the renderer / geometry branches fire only on role=annotation / the kiroAnnotation=1 marker, no shipped golden uses either -- regen_examples.py --check byte-identical; 4 new G tests + draw/geometry suite green.
…nt discovery, xdist (1.13.0) Verification (worktree, PYTHONPATH=src): - pytest FEAT-001 files + test_ruleset_sync + test_cli_discovery: 63 passed - pytest -q -n auto (full): 2696 passed, 10 skipped, 1 xfailed (16m12s) - rule_engine.cli --all --fail-on error,critical: rc 0 - rule_engine.raster_gate: rc 0 (17 rasters) - rule_engine.snapshot_gate --root examples --strict: OK - scripts/regen_examples.py --check: fresh (no golden changed)
…ified text metrics, caption overlap, border clearance, collinear kinds (1.13.0) Items 6, 7 (lint), 9, 16, 17, 18, 19 of the 1.13.0 engine review. Goldens regenerated (HA landscapes x4 + azure HA summary, aws/01, aws/03, cross-cloud/01); reasons in CHANGELOG. Solver objective keeps the frozen pre-1.13.0 caption-band / 4px border-ride models (no golden metric moved). Verified: targeted FEAT-002 tests + test_ruleset_sync pass; lint --all --fail-on error,critical rc 0; verify_icon --all --strict rc 0; raster_gate rc 0 (17); snapshot_gate --root examples --strict OK; regen_examples --check fresh; determinism PYTHONHASHSEED=1/987654 --check OK on 5 changed builders; full suite -n auto: 2740 passed, 10 skipped, 1 xfailed.
…e overlays, spec membership, respect_pins, peering from containers, AWS peer kinds (1.13.0) Task items 2, 3, 5, 8, 13, 14, 15 (FEAT-003). Goldens regenerated in the next commit. Verification (worktree, PYTHONPATH=src, shared venv): - FEAT-003 tests + test_ruleset_sync: 39 passed - full suite -n auto: 2772 passed, 10 skipped, 1 xfailed - rule_engine.cli --all --fail-on error,critical: rc 0 (27 WARNINGs, identical to HEAD~) - verify_icon --all --strict: rc 0; raster_gate: 17 rasters OK; snapshot_gate --root examples --strict: OK - regen_examples.py --check: fresh (with the regenerated goldens)
Marker x pins on long unpinned marker edges (15 goldens), overlay ring cells on the standby/supporting nodes (HA landscapes, kubernetes landscape/summary). regen_examples.py --check fresh; lint --all identical findings; aws/02 and kubernetes/01 landscapes byte-identical under PYTHONHASHSEED=1/987654.
… bands, post-layout lint check (1.13.0) - caption_routing: _ADMISSIBLE_BUDGET (250000 per planning run) and _OVERLOADED_BUDGET (2000 per overloaded-face search while the corner-exit fallback is off); truncated rankings are never cached. - crossing_repair: result-identical speed-ups (_best_group precomputed per-candidate terms + level-wise pruning, per-run ranking memo). - diagram_layout: _FINISH_NODE_GATE removed; every diagram is planned and the plan is kept only when rule_violations errors, crossings and container-overlap findings do not grow (retry crossing+collinear alone). - pipeline: Variant A bottom-band step, distinct contacts per face (2b6); geometry.rule_violations ranks *-identical exit/entry findings as errors. - layout/postcheck.py: lint-only ERROR rules on the finished layout() result. Verification: new/updated tests pass; regen_examples --check fresh (no golden changed); lint --all --fail-on error,critical rc 0 (27 WARNINGs, unchanged); verify_icon --strict, raster_gate, snapshot_gate rc 0; PYTHONHASHSEED 1/987654 byte-identical (aws/02 landscape + summary, SEP 02, SEP 06); full suite -n auto: 2796 passed, 10 skipped, 1 xfailed.
…regression bounds; docs (1.13.0) Task items 4 and 24 (FEAT-005). - tests/fixtures/sep/scripts: current SEPARCH builders copied verbatim (cmp-identical); old coordinate-free 03/09 git-moved to build_03/09_coordfree.py, pre-sync 05 kept as build_05_coordfree.py, new build_10_coordfree.py; they write under repro/. Mappings (incl. local/) synced, FEAT-003 container kinds kept; custom + opensource-colorized assets added; triples 01-13 + 01b rebuilt. - diagram_layout.build_diagram filters explicit legend_lines through _drop_undrawn_boundary_lines (inner line dropped when nothing is nested and kinds are declared); STANDARD_LEGEND_LINES leaves __all__. - Stale tests repointed (ws8, ws11, ws5, ws10, ordering_ws4, variant_a) with recorded bounds; new tests/test_sep_levers_1_13_0_fix.py (L1/L2/L5/L6). - docs/REVIEW.md, steering (diagram-standards, diagram-lint, asset-packs), CHANGELOG 1.13.0 header qualified + two Fixed entries. Verification (worktree, PYTHONPATH=src, shared venv): - pytest -n auto: 2807 passed, 10 skipped, 1 xfailed (10m37s) - pytest -n auto -m sep: 30 passed - rule_engine.cli --all --fail-on error,critical: rc 0 (findings identical to HEAD) - regen_examples.py --check: fresh - grep 'documented gap|0 own-caption cuts on the corpus and SEP' steering+REVIEW: none
- postcheck reports container-padding caption-outside-container (item 17 landscape ERROR) in layout_warnings and sets degraded (item 12 seam). - build_diagram(pinned_ids=...): layout()-routed edges no longer logged as "rerouted hand-pinned edges"; rule-engine-draw and the layout()-based golden builders pass () (item 15 seam). Output byte-identical. - canvas_metrics names its Xraw column. - L5 lever test comment: collinear pairs are a routing limit, not a skipped pass. Verification (worktree, PYTHONPATH=src): - pytest -q -n auto: 2809 passed, 10 skipped, 1 xfailed (9m01s) - rule_engine.cli --all --fail-on error,critical: rc 0, output identical to HEAD~ - verify_icon --all --strict rc 0; raster_gate rc 0 (17); snapshot_gate --root examples --strict rc 0; regen_examples.py --check: fresh - reconcile vs SEPARCH snapshots 2026-10-31_2044 / 2026-10-07_2118 with 03/09/10/13: rc 0/1 everywhere, never 3
- docs (item 24): drop the remaining "0 own-caption cuts on the corpus and SEP" claim from diagram-lint.md (edge-crosses-label), REVIEW.md D35 and the diagram-standards Variant A paragraph; point at the G15 measured numbers. - test_ws8 (item 4): test_sep_collinear_pairs_stay_within_measured_bounds, parametrized; synced 03 (3 pairs) and synced 13 (2 pairs) bounded at the measured residual (WS8 moves nothing: every candidate keeps the pair count or adds a crossing); coordinate-free 03 still 0, synced 08 e5/e9-a. - ws11 SEP helper builds 01/04 in a temp copy of the fixture root. - asset_paths_guard: --workspace-root (alias --repo-root); snapshot_gate help names the gates that accept it (item 22). - _drop_undrawn_boundary_lines logs the dropped Inner boundary line; docstring says only that line is filtered. - CHANGELOG 1.13.0: reconcile/check-snapshot documented divergence (item 1), build_diagram membership opt-in (item 13), grey primary account accepted by container-style (item 14), the above. Verification (worktree, PYTHONPATH=src, PYTHONDONTWRITEBYTECODE=1): - pytest -q -n auto: 2815 passed, 10 skipped, 1 xfailed (11m51s) - rule_engine.cli --all --fail-on error,critical: rc 0 (WARNINGs only) - verify_icon --all --strict, raster_gate (17), asset_paths_guard, snapshot_gate: rc 0 - scripts/regen_examples.py --check: fresh - reconcile 2026-10-31_2044 vs 03/09/10/13: rc 1 (9 skipped files); 2026-10-07_2118 vs 03: rc 0 (11 skipped files)
…led edges (1.13.0 phase 3 item 1) Tests: pytest tests/test_lint_truth_1_13_0_phase3.py tests/test_collinear_kinds_1_13_0_fix.py: pass
…s supporting (phase 3 item 1) Tests: cli --all rc=0; regen_examples.py regenerated kubernetes only
…ough its own source (phase 3 item 2) Tests: pytest tests/test_lint_truth_1_13_0_phase3.py pass; cli --all rc=0
…m Flow/edges (phase 3 item 4) Tests: pytest tests/test_lint_truth_1_13_0_phase3.py 9 pass; cli --all rc=0, no new findings
…er-edge pin guard (phase 3 item 3); docs + changelog for P0 - respect_pins=True keeps a legal pin verbatim (geometry.pin_is_legal, orthogonalise_route(stretch=False)); default flip to True measured and reverted (froze layout() seeds: SEP 08 collinear 1->6), see CHANGELOG Notes - warning says hand-pinned only for named pinned_ids; repair oracle and relayout pass pinned_ids=() - per-edge no-worse guard (_revert_worse_reroutes, _pin_guard_score, cap 12) - SEP 10 L6 bound accepts the new supporting-data-edges ERROR (fixture defect) Tests: full pytest -n auto: 2831 passed, 1 failed (l6 sep10 bound, fixed and re-run: pass), 10 skipped, 1 xfailed; regen_examples.py --check fresh; cli --all --fail-on error,critical rc=0 (34 WARNING lines); steering/changelog tests 342 passed
…rt filled for named pins (phase 3 P0 review) Ran: pytest -n auto full: 2785 pass, 4 fail (3 pre-existing test_annotation_block 'lambda' unmapped, 1 sep landscape fixture fixed with pinned_ids=()); targeted ws8_collinear + pins + respect_pins: 23 pass; regen_examples --check: fresh.
…eview B1)
respect_pins=None (the default) now keeps a legal pin's approach lane whether
the pins are inferred from Edge.points (pinned_ids is None) OR named explicitly
(pinned_ids=("e1",)). Before, naming a hand pin dropped it from keep, so a
named legal 20px pin was stretched to STAIR_STEP -- naming a pin gave LESS
protection than leaving it inferred, contradicting the CHANGELOG/docstring
promise. Only respect_pins=False, or respect_pins=None with pinned_ids=()
(engine/solver seeds), opts out. respect_pins=True and every builder's
pinned_ids=() path are byte-identical to before (verified exhaustively across
all nine respect_pins x pinned_ids combinations: exactly one combo changed,
the named-pin case the review asked for), so all goldens are provably
unchanged. Stale docstring/comment that still said "default stays False"
corrected to describe the None default.
Added test_named_legal_pin_keeps_its_lane_by_default (geometry, not just the
log message).
Ran: pytest -n auto test_build_diagram_pins_1_13_0_phase3 (10 pass incl new) +
test_respect_pins_1_13_0_fix (14 total); test_ws8_collinear + ws11 + variant_a
(29 pass, 6m16s); test_sep_levers_1_13_0_fix (4 pass, 3m42s); lint_truth +
collinear_kinds (23 pass). regen_examples --check: only STALE are data-URI vs
file-path icon embedding (assets not fetched in this worktree), no geometry diffs.
…lt (phase 3 P0 review B1)
Blocker B1: the 1.13.0 respect_pins=None default (keep already-legal pins
verbatim) changed a shipped golden. build_cross_cloud_example.py and
build_generic_example.py call build_diagram with edges whose points are
hand-authored route SEEDS to normalise, passing neither pinned_ids nor
respect_pins. Under the new default those seeds were kept, drifting
examples/cross-cloud/01-cross-cloud-composition.drawio (e6 entryY
0.769 -> 0.513, waypoints 10,670 -> 20,670).
Both generators now pass pinned_ids=() like every other shipped generator,
restoring the committed geometry. Verified: regenerated e6 == committed
(entryY 0.7692..., waypoints 10,670/10,540); full geometry/structure diff
of both regenerated goldens vs committed is empty once the icon-embedding
payload (data-URI vs file-path, a function of unfetched vendor assets) is
normalised. generic --check: OK. cross-cloud --check STALE only on the
icon-embedding payload (pre-existing, assets unfetched), geometry identical.
Also corrected stale prose: diagram-standards.md item-3 ("default stays
respect_pins=False" -> respect_pins=None, pinned_ids=() to opt out) and the
stale docstring in test_build_diagram_pins_1_13_0_phase3.py. CHANGELOG note
added under ## [1.13.0] ### Notes.
Tests run (.venv, PYTHONPATH=src, pytest -n auto):
- test_build_diagram_pins_1_13_0_phase3.py, test_respect_pins_1_13_0_fix.py,
test_lint_truth_1_13_0_phase3.py: 23 passed.
- Full fast suite (-m "not slow and not sep"): 2751 passed, 28 skipped,
1 xfailed, 3 failed. The 3 failures (test_annotation_block_1_13_0.py,
"no icon for role 'lambda' under provider 'aws'") are PRE-EXISTING on the
clean HEAD (confirmed via git stash) and are an unfetched-mappings/assets
environment issue, not a regression from this change.
…ts, islands, legs helper (phase 3 item 5) - _validate_spec duplicate-slot key includes leaf container (F8) - _route_container_edges separates identical boundary-edge contacts (F9) - graph_components resolves a container endpoint to its members (F18) - diagram_layout.peering_legs helper for pinned labelled legs (F19/G23) Ran: pytest tests/test_peering_container_endpoint_1_13_0_phase3.py (4), _fix (6), test_g23_peering_node_1_11_0 (incl), test_ws5_return_flows — all pass; cli --all rc 0; regen_examples.py --check fresh (goldens byte-identical).
…ase 3 item 6 F1)
check_edge_crosses_container_label scopes the group exemption to a member's
own final approach leg, so a run along the caption is flagged (06 e8, 07 e7).
Ran: pytest test_flow_group_obstacles_1_13_0_{phase3,fix} + caption/group/
container_label subset (152) pass; cli --all rc 0; regen --check fresh.
…n-aware word count (phase 3 item 7) - build_diagram(node_caption_wrap=N) wraps an overhanging node caption (7b) - caption-outside-container surfaces on flow too (7c, confirmed) - node-label-length ignores bare punctuation tokens +/&// (7d) - 7a engine leaf growth deferred (Notes): cascades into sibling/overlap Ran: pytest test_caption_sizing_1_13_0_phase3 (7) + node_label/caption subset (116) pass; cli --all rc 0; regen --check fresh.
…ottom drop (phase 3 item 10, partial)
_find_move offers the corner-contact exit for a non-overloaded lone bottom
drop too, kept under the no-worse guard. Goldens byte-identical (own total 5);
SEP coordfree unchanged (09 e10/e2 blocked by no-worse / not a lone drop).
Ran: pytest test_variant_a_bottom_bands_1_13_0_{phase3,fix} + caption subset
(15) pass; regen --check fresh; measured own cuts recorded in CHANGELOG.
…ase 3 item 11 F16) _same_face_trunk exempts two same-source/same-target legs on one face held apart by exit/entry-thirds (they never meet). 08 L5 col 3->1; goldens col 0. Hinge-move (E9) deferred. Ran: pytest test_collinear_trunk_1_13_0_phase3 (3) + collinear_kinds + sync (14) pass; 08 L5 measured col=1; cli --all rc 0; regen --check fresh.
…e, --file scoping (phase 3 item 13) - _canvas_height covers edge waypoints + per-line caption bands (13a) - _fit_size_budget re-exports under the byte budget (13b, stub-tested) - check_rasters(files=)/--file scopes the gate to one diagram (13c) Ran: pytest test_raster_export_1_13_0_phase3 (5) + raster/export subset (114) pass; raster gate rc 0; regen --check fresh.
…se 3 item 12a F20) _coverage_meta_of reads coverage_class/scope/reconcile_exempt from the companion; exempt roles with a justification drop out of omissions (recorded on ReconcileReport.exempted). Missing reason is a fail-closed usage error. 12b edge-truth checks deferred (Notes). Ran: pytest test_reconcile_scope_1_13_0_phase3 (4) + raw_snapshot (17) pass.
… document Legend offset (phase 3 item 14) - Callout dataclass + callout_cell (callouts= arg), text cell not a node - _render_node injects scope=/external=1 into a custom renderer's style - document-mode Legend sized from the index, not the full step list - companion_path / placed= membership default deferred (Notes) Ran: pytest test_build_diagram_ergonomics_1_13_0_phase3 (4) + build_diagram subset (53) pass; regen --check fresh.
…scope exemption, container_style helper (phase 3 item 15) - 15c: classify_container_kinds trusts an explicit kind= style token over the signature match; Boundary.container_kind + _boundary_style emit it (default '' -> byte-identical). - 15b: check_scope_placement exempts a VPC-scoped role inside an unambiguous peer_account/external_cloud leaf from vpc-scoped-outside-vpc; primary account still flags (fixes the broken uncommitted signature-based attempt). - 15d: neutral edge_lane container kind in constants + all 6 provider mappings (dashPattern=1 4, no signature collision) + provider-profiles.md. - 15a: resolve_container reads the bundled mapping (not a stale workspace copy, confirmed); add public diagram_layout.container_style(kind, provider) helper. - 15e token carry-over landed under item 14; vpc-nesting-under-peer_account deferred (Notes). Tests: tests/test_scope_containers_1_13_0_phase3.py (22), tests/test_mappings.py. Ran: pytest -n auto on scope/containers/mappings/geometric_containment/peer suites (83 passed); cli --all rc 0 (27 WARNINGs, unchanged).
- _assign_nodes_to_leaves keeps an external=True node assigned to its container when the container kind is an external band kind (peer_account/external_cloud/ edge_lane), so the band has children to size around (fixes the F14 SpecError 'has no children to size around'); a cloud container on an external node is still ignored. - new _EXTERNAL_BAND_KINDS in layout.base. - draw_cli._container_style emits the item-15c kind= token for a leaf band kind, so an engine-drawn peer_account classifies unambiguously and an external member inside it no longer trips external-in-account. - row_of (G19) deferred (Notes); exit_face on a band member is accepted. Additive: no shipped golden declares an external band -> byte-identical. Tests: tests/test_external_band_1_13_0_phase3.py (7). Ran: pytest -n auto on l2/external_band/scope_containers/draw/postcheck/place/peering (465 passed, 1 pre-existing annotation-test failure unrelated to this change).
- variants._is_return now also qualifies a reply whose target is a full row below the source and not to its right, so _route_return_bus offers the out-of-row bus lane for that case (SEP 05 e7's five crossings), not only a same-row leftward reply. - bus stays variant-scored (face_swap): taken only on strictly fewer crossings with no new finding, so goldens are byte-identical (regen STALE set and cli --all WARNING count unchanged, verified clean-HEAD vs change). - 8b (feedback-arc-set ordering) deferred (Notes): ordering reorders slots within a lane, never across lanes, which a request/reply cycle needs; a FAS pass is a structural change with golden-shift risk, the plan's named cut. Steering: diagram-standards.md (return-bus paragraph). Tests: tests/test_return_bus_lower_row_1_13_0_phase3.py (5). Ran: pytest -n auto on ws5/return-bus/variant/golden suites (113 passed); cli --all rc 0, 26 WARNING lines (unchanged vs clean HEAD).
resolve_icon falls back to the containers table for any CONTAINER_KINDS entry
and requires a valid #RRGGBB brand_hex; the generic edge_lane had only a style,
so the resolve-is-total and brand-color property tests failed on
('generic','edge_lane'). Add brand_hex '#666666' (its grayscale caption colour),
matching every other provider's edge_lane entry.
Ran: pytest -n auto tests/test_icon_resolver_properties.py
tests/test_icon_resolver_brand_color_property.py tests/test_mappings.py (49
passed).
…ers off container captions (phase 3 item 16) - diagram_layout._overlay_labelled: skip -standby suffix when the label already names the standby channel (hyphenated or spaced word), so 'RDS standby' stays (10 feedback). layout.pipeline._degraded_finish re-runs the oracle on the FINAL normalised candidate so layout_warnings/degraded reflect the returned geometry (12 container-padding noise gone). geometry.check_marker_on_container_caption + linter marker-on-container-caption (WARNING; ERROR on network/component/flow, lint-only); resolve_marker_collisions phase 4 slides a marker off a container caption and phases 2-3 never slide onto one. oci/01 golden refreshed (e3 marker slid off the region caption, drawio+png re-exported). external_cloud leaf caption-strip resize NOT shipped (moved SEP 09 geometry; CHANGELOG Notes). - Steering: diagram-lint.md rule row+detail, diagram-standards.md. Tests: tests/test_rendering_residuals_1_13_0_phase3.py (10 passed). - Verified: pytest item16 + overlay + marker + linter + ruleset_sync + init_workspace + golden_examples (all pass); cli --all rc 0, 34 WARNINGs unchanged; SEP L2/09 exit-face lever still passes.
…gion/AZ/Subnet, edge-label offenders (phase 3 item 17) - geometry.text_label_boxes + check_text_label_on_container_border: a NON-numeric text edge label (a peering leg label, a descriptive label) riding a Boundary border is reported under marker-on-container-border (F3, 08 peering leg on the account border); rectangle-vs-line test over the centred label box. - geometry.check_legend_boundary_coverage extended (F5): a Legend naming a Region / Availability Zone / Subnet not among the drawn container KINDS is flagged; precision-first, linter passes drawn_kinds classified via classify_container_kinds, None for C4/unrecognised -> structural-only. - edge-label already names offenders and has no peering exemption (F19) - covered by explicit tests. - Steering: diagram-lint.md edge-label/marker-on-container-border/ legend-boundary-coverage rows + detail. Tests: tests/test_lint_residuals_1_13_0_phase3.py (10 passed). - Verified: pytest item17 + ruleset_sync + linter + marker-on-container-border + ws11_legend_coverage + rule_offenders (110 passed); cli --all rc 0, 34 WARNINGs unchanged.
…wn-container peering label (phase 3 item 18, item 17 refinement) - diagram_layout.Edge.unconfirmed: a dashed edge carrying overlay=unconfirmed for an inventory-unbacked relationship (11 SQS consumer, 09 e21); numeric marker value unchanged so flow-marker/flow-legend checks still see a plain number. overlay_legend_lines(nodes, edges=()) scans edges for the standard line. - _OVERLAY_SHAPE/OVERLAY_LEGEND_LINES gain no-known-caller (03 strapi-alb cluster); like every overlay-marked node it is in geo.overlay_nodes so graph-islands excludes and node-connectivity accepts it. - item 17 refinement: check_text_label_on_container_border exempts a container the edge legitimately enters/leaves (endpoint == container, or a node inside it), so a materialised peering leg's label near the VPC face it exits is not a false ride (keeps test_g23_peering_node clean); only an unrelated frame fires. - Steering: diagram-standards.md Overlay Vocabulary + unconfirmed-edge/ no-known-caller paragraphs. Tests: tests/test_overlay_terms_1_13_0_phase3.py, updated tests/test_lint_residuals_1_13_0_phase3.py. - Verified: pytest item18 + g23_peering + peering_container + lint_residuals + ruleset_sync + init_workspace + overlay_ring (all pass); cli --all rc 0, 34 WARNINGs unchanged.
…cident nodes -> SpecError (phase 3 item 19)
- routers.CoincidentNodesError (subtype of UnclassifiableEdgeError): classify_edge
raises it for dx==dy==0 naming both nodes; layout() catches and re-raises a
SpecError ('occupy the same position ... give one a different slot or sub') -
the SEP 08 IGW-at-slot-9 case that was an internal UnclassifiableEdgeError.
- advise_cli.entry_face_suggestions: for each end-to-end-merge collinear pair on
the built geometry, propose the alternative legal entry face (top<->left) for
the later-id edge (SEP 09 e12/e26); text + JSON, empty when nothing merges.
- Up-front (lane,region,slot,sub) key NOT tightened (would relax same-slot/
different-sub rejection goldens rely on); late catch is the robust fix. Notes.
- Tests: tests/test_advise_specerror_1_13_0_phase3.py (5 passed).
- Verified: pytest item19 + ordering_ws4 + flow_groups (42 passed); cli --all
rc 0, 34 WARNINGs unchanged.
…se 3 verify) Root cause: crossing_repair._accept guarded collinear only with a COUNT cap, so a crossing-fixing move could trade one plan-start collinear pair for a DIFFERENT one while the total stayed flat. Item-11's trunk-exemption widening changed caption_routing._score.collinear and thus the move chosen on the SEP 03 coordinate-free landscape: e13 (apigw->rds) landed collinear with e4 (lambda->sqs) as an end-to-end-merge the later collinear-shift pass could not undo (base 1.13.0 produced 0 collinear pairs on that WS8 acceptance fixture). Fix: crossing_repair._end_to_end_merges captures the end-to-end-merge pairs at each move's start (_Base.merges); _accept rejects any trial that introduces a merge absent from that set (re-using one already there stays allowed). A genuine end-to-end merge is two unrelated edges, never a fan-out/fan-in trunk, so the guard does NOT weaken item 11's lint exemption; trunk / pass-through pairs the collinear-shift pass clears are untouched. Measured (real SEP builders, base 1.13.0 -> HEAD): SEP 08 collinear 8 -> 0; SEP 03 and SEP 13 unchanged at 0. WS8 fixtures: 03 coordfree 1 -> 0, 08 1 -> 0 (within their test_ws8_collinear bounds). The AWS/Azure/GCP/OCI HA landscape goldens are regenerated separately (next commit). Also: re-export CoincidentNodesError from rule_engine.layout_engine (item 19 added it in layout.routers but not the facade); repair two pre-existing 1.13.0 tests broken by earlier phase-3 field changes (test_layout_engine coincident-nodes message = item 19's CoincidentNodesError; test_container_caption_wrap Edge(...) by keyword after item 18 inserted the unconfirmed field). Test-only, no behaviour change. Ran: pytest test_crossing_merge_guard/test_ws8_collinear/test_collinear_trunk/ test_collinear_kinds/test_layout_engine/test_container_caption_wrap pass; full suite 2956 passed/10 skipped/1 xfailed under PYTHONHASHSEED=7.
The four HA-landscape goldens (aws/azure/gcp/oci) and oci/01 were stale against the HEAD engine: a uniform l13 wafedge->cdn reroute (exitY/entryY 0.2564 -> 0.5128) from the phase-3 engine changes, plus the aws HA l10/l14 reroute from the end-to-end-merge guard. Regenerated via scripts/regen_examples.py with the vendor asset packs present (no data-URI vs file-path churn). All stay lint-eligible (rule-engine-lint --all --fail-on error,critical rc 0) and byte-deterministic (regen_examples.py --check fresh; identical drawio hashes under PYTHONHASHSEED 0 / 12345 / 7). CHANGELOG: two phase-3-verify Fixed entries under [1.13.0] (the end-to-end-merge guard; CoincidentNodesError re-export + the two repaired tests, with the three test_annotation_block lambda-unresolved failures noted as a pre-existing unfetched- assets environment condition, not a regression). Ran: regen_examples.py (ok), regen_examples.py --check (fresh), rule-engine-lint --all --fail-on error,critical (rc 0).
…n the final engine Verbatim copy of the CURRENT SEPARCH builders (scripts/ + .build-tools/ mirror the SEPARCH layout), their NN-*.diagram.md companions, and SEP's mappings/ (incl. the local/ role+icon overlays that resolve the SEP-specific roles dial/argocd/neptune/...) under tests/fixtures/sep/separch/. test_sep_regression_1_13_0_phase3_verify builds each of the 14 builders in a temp dir mirroring the SEPARCH layout (so every Path(__file__) / ROOT-relative write lands in the temp, never in SEPARCH), drops the committed companion beside the .drawio, and asserts the pair is lint-eligible (0 ERROR/CRITICAL) under the stricter final 1.13.0 rules. Result: all 13 SEP diagrams (14 builders) stay lint-eligible. Assets are optional (the engine embeds a file-path icon ref when the packs are absent, still eligible); the per-provider mapping cache is cleared around each build so the SEP local overlays are read under the temp CWD (test isolation). Ran: pytest test_sep_regression (14 passed, parallel and serial); also with the workspace assets/ symlink absent (3 spot builders pass).
Two additive REVIEW.md gap entries summarising the phase-3 engine work (lint-truth, routing, container and overlay residuals from the SEP phase-2 feedback), each pointing at the per-item CHANGELOG [1.13.0] detail.
…ression fixture The SEP 01-13 regression fixture (637583e) mirrors the SEPARCH layout, where six of the 14 builders live under .build-tools/ (01/02/03/04/09/10). The root .gitignore's '.build-tools/' pattern silently dropped them from that commit, so test_sep_regression_1_13_0_phase3_verify failed with FileNotFoundError for those six diagrams. Copy the canonical SEPARCH builders into the fixture and add a scoped .gitignore negation so they are tracked (the root .build-tools/ stays ignored). All 14 SEP cases now lint-eligible on the final 1.13.0 engine.
…t the pack-only lambda alias test_annotation_block_1_13_0.py (feedback G) declared its incidental service node as role='lambda'. That alias resolves only from the fetched AWS vendor pack, so the asset-less CI 'test' job failed three cases with DrawError: no icon for role 'lambda'. Switch the node to 'serverless_fn' (present in mappings/roles.yaml and the committed icon-index.json), so the tests resolve with or without the packs. The feature under test (role='annotation' as a placed edge endpoint) is unchanged. CHANGELOG note updated to reflect the fix.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Release 1.13.0 — engine routing, caption & peer-sizing levers for dense SEP accounts.
Full release notes: see the
## [1.13.0]section of CHANGELOG.md in this PR (thesection is too long for a PR body — GitHub caps it at 65536 chars). The GitHub
Release notes are extracted from that CHANGELOG section by release.yml on the tag.
This PR carries the 1.13.0 release commit plus a test-fixture fix: the six
.build-tools/SEP builders (01/02/03/04/09/10) that the root.gitignoresilently dropped from the regression fixture are now committed, so
test_sep_regression_1_13_0_phase3_verifypasses for all 14 SEP diagrams.