WSL install: add no-start mode and host port allocation - #1180

Scope

WSL distributions share one network namespace. This PR lets a new distro install the selected stack without activating services and gives provisioning, clients and acceptance one complete host allocation. A selective install now refuses a requested port change for an unselected owner; later consumers cannot silently use an allocation different from the installed snapshot.

The decision-preservation guard now accepts append-only additions while requiring every byte from the frozen8053f14e decisions to remain an unchanged prefix. This fixes the CI failure caused by #1158's legitimate append. Original digests are retained with their original byte lengths; source exports still need no Git history. The delegated wave5 negative control now changes an original byte instead of calling an append a rewrite.

• Base: 8053f14e7e0817d2502dff2bb284bfb7e1b0152f; lane: lane:foundation.
• Owned paths: install-plan artifact, host/client inputs and tests, plus authorized runtime/research/dashboard readers, citation repins, recipe/checklist/handbook consumers and mechanical evidence registration. Handbook receipt grant186741 preserves existing history and appends regenerated bindings.
• This is the existing non-draft PR. The implementation fold was one normal commit on c231aebcc29d70f88aeac0ae92385e25d1b6c241, authorized by196445/196662 and confirmed by197743. CC200867 subsequently authorized the landing rebase onto main 24f3e7827b47c4bb95c63096ea2accb7800aafda, preserving the reviewed change.
• CC202609 authorizes one normal corrective commit on d4253c8cdf3dd69329d176c340dc3ca7a247a058 for the prefix guard, its regressions and registered evidence. Decision documents and production code are unchanged in this correction.

install.sh --no-start installs units and reloads systemd without enabling or starting them. Dagu uses its documented no-service mode. Docker package activation is masked before package work and rootless setup is skipped, including Compose/Harbor dependencies. MinerU prepares Standard models through its standalone kit; the recipe and checklist require a normal --only mineru --host-file rerun without --no-start before activation to disable telemetry and configure managed mode. Existing enabled/running state is preserved.

Complete host port keys drive provisioning, readiness, acceptance, telemetry, gateway and client rendering through the plan's renderer. Client configuration also requires the target distro identity. Optional ports remain reserved against collisions. Grafana gRPC is required only when enabled; codebase-memory's optional UI has an explicit disabled native configuration. Both Promptfoo predicates compare the effective HTTP port, including80. The original default host rendering remains byte-identical within the recorded historical asset scope.

For example, if the committed gateway is41128 and a host file requests51128, installing only memory-owner now refuses before writing row assets. Acceptance, both public client APIs, runtime telemetry/providers and MCP consumers refuse that same disagreement. A selected-owner update followed by publication allows every consumer to use the new allocation. Client checks use the target HOME/config root; inherited XDG_CONFIG_HOME applies only when rendering for the current HOME.

Research-only installation writes the selected gateway into GPT Researcher's existing native JSON LLM_KWARGS.base_url. Later invocations retain it after the temporary installer selector disappears, as does a subsequent memory-only install. A conflicting explicit allocation is refused. Research publishes no listener ownership snapshot and creates no additional consumer-state file. The default research JSON remains byte-identical.

Prior corrections remain: source-derived custody for31 copied assets across13 portless rows; idempotent memory bind/provider migration, including49374; exact Dagu vendor-unit/drop-in preflight; Grafana operator-edit preservation; Prometheus install-time working-directory creation; and normal MinerU telemetry-before-download ordering. Publication retains staged intent, prior-snapshot comparison and advisory locking. These controls protect snapshot publication; they do not provide rollback for every earlier asset write. Recovery is documented in NO_START_PORTS.md.

The live recipe/checklist and generated handbook require complete host input, no-start provisioning, separate activation and a fresh listener-ownership check. The dated addendum supersedes historical commands without rewriting the original decisions. The frozen nine-field template, image experiment and sealed critic remain unchanged; template extension is deferred. Handbook publication uses dated attributed paraphrases and preserves source evidence/history under189041. Private originals retain name/custody only, with no digest.

Landing dependency: #1161 and #1179 overlap the handbook receipt/current output. Any later overlapping landing takes a rebase and native regeneration, preserving append-only receipt history. Never hand-merge receipt.json. The CC-ordered rebase onto 24f3e782 encountered no handbook/input conflict; handbook bytes and bindings remain unchanged. It claims no future merged-tree validation.

The landing rebase preserves all eight implementation commits. Six patches are identical; the remaining two differ only in manifests/evidence.json. One conflict was the hash/size registration for the automatically merged tests/test_provenance_digest_custody.py, refreshed with the native registration helper. The other was adjacent additions: both the main-branch Nautilus RC5 decision and this PR's WSL host-allocation addendum were retained in sorted order. No production or test code required manual conflict resolution. The final reviewed 6315ebbfb patch is identical after replay. The two subsequently identified P2s are assigned to a separate follow-up from main after this PR merges, required before NativeHorizonNY normal-mode activation.

Rebase verification at d4253c8cdf3dd69329d176c340dc3ca7a247a058: scripts/validate.py passes12,246 hashed files/71 components/4 profiles/240 receipts; all33 convergence records pass; the full24-test provenance-custody module passes with0 failures/errors/skips. These checks run detached with idle CPU priority and individual exit files. The direct old-tip/new-tip comparison also contains main's automatically merged provenance-test additions; the PR's test patch is unchanged in the range-diff.

The co-op published that rebase from a native WSL shell with the full gate passing at10:59:39Z–11:00:56Z. Earlier detached-service attempts omitted required WSL context or were interrupted; their reports remain retained and are not successful-gate evidence. Subsequent long checks use detached scopes that inherit the native caller context without manually setting identity selectors.

SOTA sources

• Frozen guard at d4253c8c, test_install_plan_review_wave3.py:225–242, original client decision at8053f14e, and append in7dcbe9a58/#1158: original content remains a prefix; appending is distinct from rewriting.
• CPython v3.13.0, stdtypes.rst:981,2547 and hashlib.rst:106–116: byte-sequence slicing and SHA-256 support the fixed-length original-prefix check. No baseline digest is repinned to current main.
• systemd v259, systemctl(1), man/systemctl.xml:880–927: enable creates installation links; --now also starts. Mask semantics:1129 governs package-triggered activation.
• systemd v259, systemd.unit(5), man/systemd.unit.xml:2224–2255: install-time link semantics.
• Dagu v2.18.2, scripts/installer.sh:113, unit generation:1666–1698, and service branch:2180: native SERVICE_MODE, unit output and no-service installation.
• Recipe F8 at8053f14e7, linux-wsl2-new-distro.md:1047 and Listeners and ports, linux-wsl2.md:253: host inputs and WSL's shared network namespace.
• Existing selective merge at c231aebc, observability_config.py:274–277 and research gateway consumer:45–47: reference implementations for the two selective-install corrections.
• GPT Researcher0957c301, config.py:63–76 and 158–170: native JSON configuration and LLM_KWARGS consumption.
• MinerU c221cc41, models.py:86–116,171–205, telemetry.py:39–57, and kit/main.py:42–54: standalone preparation and daemon-dependent telemetry/managed configuration.
• ai-memory2.6.6 at04a9e8f5, init.rs:15,27–35,99–106 and default TOML: fresh init writes the embedded template and appends auth. Its active defaults contain no embedding provider. The fixture uses synthetic auth and is not a native-init claim.
• moby8af9fe3a, dockerd-rootless-setuptool.sh:334–390: native unit form and setup start/enable calls.
• Grafana v13.2.3, setting_grpc.go:113 and codebase-memory-mcp v0.11.0, src/ui/config.c:124–157: optional listener defaults and native disable configuration.
• CPython v3.13.16, fcntl.rst:176–187: advisory locking between cooperating snapshot publishers.
• Further pinned source mechanisms and prior correction evidence are retained in SOURCES.md, NO_START_PORTS.md and the per-wave receipts.

Evidence-class table

Local commands run

All heavy checks use systemd-run --user --scope -p CPUWeight=idle --quiet --. The temporary runner invokes native unittest suites and emits only counts/test IDs/code locators; it adds no skips.

Latest prefix correction: the full wave3/wave5 modules plus tests.test_install_plan_frozen_prefix pass35 tests with0 failures/errors/skips. Its fail-first run against unchanged d4253c8 ran6 tests with5 assertion failures and0 errors/skips. Both current documents were independently compared with main9d5bc9f9 and the original8053 blobs: the client decision has828 appended bytes, and both original blobs remain intact prefixes. check_plan.py passes84/198/113. Required pre-commit validation passes12,247 hashed files/71 components/4 profiles/240 receipts; all33 convergence records pass.

The following larger wave7 runs predate the landing rebase and prefix-only correction; their recorded scopes are retained:

python3 -B /tmp/install-nostart-safe-tests-velo.py tests.test_install_plan_no_start_ports tests.test_install_plan_review_wave3 tests.test_install_plan_review_wave5 tests.test_install_plan_review_wave6 tests.test_install_plan_review_wave7 tests.test_runtime_profiles.StrictHostConsumerTests tests.test_harbor_currency_contracts tests.test_research_entry tests.test_wsl_new_distro_recipe tests.test_new_wsl_handbook tests.test_new_wsl_handbook_privacy
exit0:328 tests,0 failures,0 errors,0 skips

python3 -B evidence/artifacts/new-wsl-install-plan-20261002/check_plan.py
exit0:84 rows,198 commands,113 acceptance entries

python3 scripts/validate.py
exit0:12,160 hashed files,71 components,4 profiles,240 receipts

python3 -B scripts/validate_convergence.py --all-recorded
exit0:all33 records valid

git diff --check
exit0

The nine new wave7 tests all pass. Initial unchanged-c231aebc evidence is4 tests/4 failures/0 errors. A separate corrected replay of the exact prior public function ASTs gives1 test/2 expected assertion failures/0 errors. It is an isolated function replay, not a full baseline checkout. MinerU's documentation test gives1 test/2 failing subcases before correction. Incomplete-input draft replay/XDG errors are explicitly excluded from fail-first evidence. The corrected XDG test is positive regression coverage.

Repository source/AST/grep inspection covers direct and indirect consumers in strict-host-consumer-audit.json. Graph transport was unavailable; no graph-completeness claim is made. Candidate checks intentionally validate proposed host allocations; runtime readers validate committed state. runtime_profiles.py needs no further edit in this fold.

Pre-edit install/accept citation inventory retained78 hits/37 files. The research command changes on its existing shell line; install line count and accept bytes remain fixed. All8 current anchors remain valid. The renderer's current receiver selector is repinned to523–536; runtime locators43 and1027–1034 and research wrapper line27 remain fixed.

Evidence registration uses the native register/normalize functions. Compared with8053,59 existing entries change only hash/size,21 entries are added and none are removed; top-level/extra metadata and five frozen inputs remain intact. A protected-prefix scan of93 inputs covering changed/new files and the PR body reports0 matches at lengths7/8/12/16/64, using previously published metadata in memory only. No private original was opened or hashed.

Prior validation remains historical: wave6 passed112 focused tests and111 recipe/documentation cases (two overlap); wave5 passed198. Full-shard results and baseline controls remain in review-wave-4-receipt.json, including baseline-matched failures and native-environment skips. They are not new full-shard or native acceptance claims. No production CI skip was added.

Decision record

frozen-prefix-receipt.json records the current CI correction, unchanged original hashes/lengths, before/after controls, pinned sources and execution boundaries. The prior records below retain their historical scopes.

NO_START_PORTS.md, review-wave-7-receipt.json, review-wave-6-receipt.json, preceding receipts, the host-allocation addendum, and the consumer/citation audits. Current review sources are filename-only: review_output.json, MICRO-delta-c231aebc.md, GPT-VERDICT-c231aebcc.md.

Host evidence

native-agent-stack-51 runs install.sh --no-start --only memory-owner with the complete NativeHorizonNY host file, then verifies that the unit is disabled and no new listener appeared. This lane has not performed that acceptance. All wave7 execution uses disposable fixtures, extracted commands and fakes; no full installer, full acceptance script, native init, service activation or model download ran.

An early broad legacy run invoked accept.sh with disposable HOME/stub curl and included a synthetic native-collector harness. That exceeded the requested local boundary and was disclosed in the retained receipts. Subsequent local checks exclude complete-script/native-daemon cases; they remain in GitHub CI.

Follow-ups and verification boundaries

• CC/GPT delta reviews apply to the newly published head. Fresh CI and merged-tree verification belong to the landing pass.
• NativeHorizonNY owns package/bootstrap, disabled-unit/listener and lifecycle/recovery acceptance. Foreground observability lifecycle remains a proposal until measured there.
• Handbook overlap requires the landing rebase and native regeneration described above. Source evidence/history remains preserved under189041; the proposed source-cleanup follow-up was cancelled.
• Frozen template extension, needs_owner handling, static --list, recovery and later UI recheck remain the previously recorded deferred work. Current controls do not claim whole-install rollback.
• Earlier Collector source-pointer mirror availability and a dedicated fail-first wording guard remain recorded evidence follow-ups.

Checklist

• No new workflow or scripts/*.py file; existing generator source-policy grant retained.
• No new required secret or paid service; no private-original digest published.
• Peer-owned files/worktrees, frozen inputs and historical decision/receipt content preserved.
• Exact non-empty SOTA section and pinned source mechanisms included.
• Required validation before the single authorized normal commit; configured no-reply identity and normal push only.