Bounded Workflow Topology Preparation: WSL2 Linux Evidence¶

This report records the required production-topology evidence matrix for ADR-0005 and issue #141. The authoritative measurements are in the adjacent JSON artifact.

• Collected: 2026-07-22T01:30:09Z (2026-07-21 local)
• Source revision: fd08ab21373c0bf1e0586e54ec0a564e3d4ce4d5
• Source tree: a886ace2a976ddb68f6e4c5535b447dd80e9c84f
• Implementation digest: 93ae4c2165b3fa1a7ffcf9ae5a80ce6dd0b794f4284660560d93239e0a6c42e4
• JSON SHA-256: 7314c9aac63bcbce5635e31c5e089760dabf06f9d6a8abc8a658bdd6313f5661
• JSON size: 32,802 bytes
• Result: all six required cases and the forced-termination cleanup control passed

Scope¶

The benchmark exercises the issue-#141 production topology adapter through final legacy detachment. Every case consumes one-shot topology generators in a fresh child process, prepares canonical V1 pages and manifests through the private SQLite spill workspace, materializes the compatibility observed_node_ids value, and removes the workspace before the parent accepts the result.

Report schema v2 records a bounded-phase checkpoint before legacy detachment and an end-to-end high-water mark after detachment. The run did not start Ray, write a Django database row, contact Kubernetes, activate a writer, or enable schema v3.

Environment¶

The branch was exported as a Git bundle and cloned into native WSL /tmp; the run did not depend on the unavailable /mnt/d mount. PYTHONPATH pointed at that clone's src, and both source snapshots recorded the same clean revision and implementation digest.

Selected resource profile¶

All cases retained the same five direct scan or primary-key-search plans. No plan used a temporary B-tree or implicit materialization.

Results¶

Memory and spill values use MiB. The bounded and end-to-end columns are distinct schema-v2 measurements even where this run recorded the same process high-water.

Interpretation¶

After retained output reached the V1 caps, bounded-phase memory stayed effectively flat while observed input grew:

• sparse tracemalloc remained 55.67 MiB and RSS ranged from 184.60 to 186.94 MiB;
• high-edge tracemalloc ranged from 101.21 to 101.22 MiB and RSS ranged from 303.23 to 303.48 MiB; and
• retained output stayed at 25,000 nodes, with either 24,999 sparse edges or the 100,000-edge cap.

External spill grew with observed input, from 5.61 to 55.94 MiB for sparse cases and from 12.13 to 121.11 MiB for high-edge cases. That is the selected trade: exact uniqueness and endpoint-reference state scales on bounded local disk rather than in complete Python node and edge collections.

The end-to-end high-water equaled the bounded-phase high-water in this run. That does not erase the compatibility boundary: the artifact records 25,000, 100,000, and 250,000 legacy observed IDs, and issue #142 still owns removing or confining that O(observed) final value.

Source, validation, and cleanup controls¶

• The source snapshots before and after the matrix were clean and identical at fd08ab21373c0bf1e0586e54ec0a564e3d4ce4d5.
• The same tree passed uv run make ci, 267 focused preparation tests with 5 expected Windows skips, 214 Compiled Graph-selected tests with 1 expected Windows skip, and 500 seeded randomized parity cases before the Linux run.
• Every normal case reported worker-context and parent-watchdog cleanup as removed; no scenario root or workspace remained.
• The forced-termination child reached an open workspace, was killed with return code -9, created no durable candidate, and had its scenario root removed by the parent.
• The benchmark's schema-v2 validator accepted the retained artifact again after the run, and the configured workspace parent was empty.

Decision and limitations¶

This matrix and the focused canonical, failure, cancellation, ownership, cleanup, and capability tests satisfy issue #141's production-topology evidence gate. They show that the spill-backed path removes complete Python node and edge validation collections without changing canonical V1 output.

This is one run on one WSL2 host and ext4 filesystem, without concurrent preparers or repetitions, and is a memory-contract decision record rather than a latency SLO. It does not satisfy issue #142's composite topology/detail lifetime, issue #79's live wire/mailbox/aggregate admission limits, or any schema-v3 activation gate.

Reproduction¶

The retained run used the following command from the clean native WSL clone:

PYTHONPATH=/tmp/django-ray-issue141-20260721181601/src \
/tmp/django-ray-issue140-venv/bin/python \
  scripts/benchmark_workflow_progress_preparation.py \
  --implementation production-topology \
  --required-scale \
  --timeout-seconds 7200 \
  --workspace-parent \
    /tmp/django-ray-issue141-20260721181601-benchmark \
  --output \
    /tmp/django-ray-issue141-20260721181601-production-topology.json