fix(release): close plan-doctor false positives and lint parity gap; prep v0.9.0

[joshuaboys]

Release review of v0.8.1 plus the unreleased work on main. CI was green and every distribution pin agreed, but three defects surfaced. Two are fixed here; the third is left to the tool being built for it.

Closes #132.

Defects fixed

plan-doctor flooded healthy plans with false warnings (CIB-005, ISS-010)

The bundled skill judged plans against a stricter private convention than plans/aps-rules.md documents. On a consumer plan with no real defects it drew roughly 385 warnings: 162 Done and 132 Proposed statuses false-flagged by W03, and all 91 module filenames flagged by W02. Because the skill bytes are embedded in the binary, a consumer could not patch it locally without desyncing .aps-managed.json.

Status now normalises through the documented alias table before W03, W04, W05 and W07 evaluate. W04 counts Done as terminal. The filename-prefix half of W02 becomes a new Info code reported once per directory with a count, and a consistently applied undocumented status such as Archived gets its own Info code rather than sharing W03.

The fix also added a general collapsing rule to the report section: a finding that hits 100% of files is house style, not signal. That addresses the cause rather than the three symptoms. The skill now agrees with its sibling aps-planning skill, which already accepted the aliases.

Release-plan lint existed in one of three CLIs (CIB-006, ISS-011)

R001 to R004 were implemented in the Rust linter only, so the bash and PowerShell linters never discovered plans/releases/v*.md at all. On this repo's own plans bash checked 42 files where Rust checked 50, and a malformed release plan passed both fallbacks silently. D-039 makes bash and PowerShell maintained peers, which supersedes REL-003's "aps lint is now the Rust CLI" rationale.

Discovery was the larger half of the defect: get_file_type needed a release branch at the Rust classifier's precedence and find_aps_files needed a releases/ clause, pinned to byte-order sorting to match Rust. Full output including --json is now byte-identical across all three CLIs.

CI could not have caught this either. The parity harness carried no release fixtures, and its findings regex matched only (E|W)[0-9]{3}, so R codes were invisible even when present. Both are fixed.

Installer manifests did not carry the new rule module

Sourcing a rule module from bin/aps without listing it in the installer manifests breaks a vendored or scaffolded bash CLI on startup. Fixed across all fourteen sites, and a new test derives the required list from bin/aps itself so the next rule module cannot repeat it.

aps release — REL-005, Rust phase

Scope grew after the release prep landed, so this PR now also carries the top-ranked Ready work item. Four subcommands: new scaffolds a record from the template, status reports modules and work-item completion, notes drafts markdown from Complete items since the previous release, and close advances Merged/Released items to Complete with a release-evidence line.

The closeout reads the prose record plus the plan tree directly, with no JSON sidecar, which is the work item's stated design constraint. Dry-run is the default and was verified to write nothing by diffing a scratch copy of plans/ afterwards. Two safety properties are worth calling out: --apply refuses a record whose Status is not Shipped/Released/Archived unless --force, so a release cannot be closed out mid-prep, and a record missing ship evidence warns that it would stamp today rather than doing so silently.

Two defects were caught by running it against this repo's real records rather than fixtures, and fixed:

• status reported an unshipped release as shipped today, defaulting the ship date to the planning date. A release tool that asserts an unshipped release shipped is worse than one that says nothing.
• The (cut) label used by the v0.8.0 and v0.8.1 records was not understood, so the cut date was reported as the ship date and the real one dropped.

REL-005 stays In Progress, not Complete. D-039 requires bash and PowerShell to carry the same surface with shared-fixture coverage before a new command surface is done, and only the Rust phase has landed. Phasing is explicitly allowed by the work item; shipping one-CLI-only is not. The changelog entry therefore sits under Unreleased rather than in the v0.9.0 section, so the release narrative's out-of-scope list stays accurate.

Plan reconciliation

aps audit went from two findings to none, and the only lint warning is gone.

• PROMPTS-001 and EXAMPLES-001 marked Complete against their real merge evidence; the prompts and examples modules close.
• cli-redesign moved Draft to In Progress with six of seven items done; D-045 recorded as decided rather than proposed.
• Roadmap overview rewritten. It still listed the MCP server and GitHub Action as Future, though both shipped in v0.7.0.
• The completed archive was filing v0.3.0 through v0.7.0 task tables under an Unreleased heading, and v0.4.0 through v0.8.1 were never rolled at all. The false heading is fixed; the roughly sixty-item backfill is left to REL-005 rather than reconstructed by hand, which would have meant inventing attribution. Tracked as ISS-012.

New backlog items

Each carries an issue-log entry:

• CIB-007 / ISS-013 — neither fallback CLI can answer aps --version, though both know their version well enough to stamp it into project config. That is the exact question the pin-mismatch warning raises.
• CIB-008 / ISS-014, ISS-015 — three residual cross-CLI lint divergences (symlink traversal, PowerShell's culture-aware sort, W020/W021 grouping order) and release rules weaker than they read (R001 accepts v0garbage.md). Both reported rather than fixed in one CLI, which would have re-opened the gap just closed.

v0.9.0 prep

Version stamped to 0.9.0 across fifteen files; the release checklist's stray-version grep comes back empty and both CLIs report 0.9.0. Release narrative, changelog section and README highlights written.

Not done, and deliberately left to a human: the tag push, cargo publish, and the Scoop hash, which needs the published SHA256SUMS. The release narrative marks those channels pending.

Verification

cargo test                 230 passed
cargo clippy -D warnings   clean
cargo fmt --check          clean
test/run.sh                all tests passed
test/cli-parity.sh         bash = Rust across 12 fixtures
markdownlint **/*.md       clean
aps lint plans             50 files, no issues (bash and Rust identical)
aps audit --no-run         no findings

PowerShell parity was verified under pwsh 7.4.6, including that the ported rules actually fire rather than silently matching nothing. The local parity run skips that leg because pwsh is not on PATH here; CI covers it and passed green on the two earlier commits.

🤖 Generated with Claude Code

https://claude.ai/code/session_01N2expUMF61CKDC5zxZUeev