@decocms/parity 0.11.16 → 0.12.0

package/AGENTS.md

@@ -46,7 +46,10 @@ Do **not** run parity for:
 
 ```ts
 {
-  verdict: { status: "pass"|"warn"|"fail", score: 0..100, critical, high, medium, low },
+  verdict: { status: "pass"|"warn"|"fail", score: 0..100, scoreVersion: 2, pagesAnalyzed, critical, high, medium, low },
+  // score v2 = 100·e^(-penaltyDensity/35), penaltyDensity = Σ severity weights ÷ pagesAnalyzed.
+  // Any FAIL verdict (critical issue or failed check) caps it at 79. It's a progress meter: it RISES as issues get fixed.
+  previousRun?: { id, timestamp, score, scoreDelta },  // trend vs last comparable run (same hosts + scoreVersion)
   topIssues: Issue[],       // LLM-ranked, deduplicated; usually 5-10 items
   issues: Issue[],          // all raw issues from every check
   checks: CheckResult[],    // one per check; .data has structured details

package/CHANGELOG.md

@@ -5,6 +5,40 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [Unreleased]
+
+## [0.12.0](https://github.com/decocms/parity/compare/v0.11.17...v0.12.0) (2026-07-27)
+
+### Known issues
+
+* **A full `parity run` (multiple viewports/sides concurrently) has been observed to hang indefinitely in at least one resource-constrained sandboxed environment**, after selector discovery's own throwaway browser launch+close already succeeded. Isolated components (`launchBrowser`, `capturePage`, `parity journey`, sequential/concurrent browser launches) all completed correctly in the same environment, so this reads as environment resource contention rather than a confirmed code defect — but it was not fully root-caused. If a run seems stuck at "Launching browser…", try `--max-viewport-concurrency 1` or a narrower `--only`/`--viewports` scope, and please report reproduction steps.
+
+### Fixed
+
+* **The health score was stuck at 0 on every real run — now it actually rises as you fix issues.** The old formula (`100 - crit·20 - high·8 - med·3 - low·1`) had no dynamic range for real workloads: 13 high issues alone zeroed it, and checks emit one issue per occurrence (per page × viewport, per robots.txt user-agent, per broken link), so mid-migration runs carry 40–120 issues. Empirical evidence: across 15 real runs of the granadobr migration, issues fell 122 → 39 (criticals 18 → 0) and the score sat at **0 the entire time**. The new formula (score v2) normalizes the severity-weighted penalty by the number of analyzed page-pairs and applies exponential decay — `round(100·e^(-density/35))` — so the same 15 runs now read **20 → 7 (regression caught) → 35**, moving with every fix. Any FAIL verdict (critical issue or failed check) caps the score at 79 so "FAIL · score 91" can't happen. Status logic (pass/warn/fail), `--fail-on`, and exit codes are unchanged. Verdicts now carry `scoreVersion: 2` and `pagesAnalyzed`.
+
+### Added
+
+* **Score trend vs previous run.** `parity run` now records `previousRun: { id, timestamp, score, scoreDelta }` in `report.json` (most recent non-partial run against the same prod/cand host pair and same scoreVersion) and surfaces the delta in the CLI summary (`score 64/100 (+23 vs run anterior)`), the HTML dashboard (chip under the health ring), and the `parity pr` Markdown comment (header + score-trend line).
+* **Root-cause grouping in issue displays.** New display-only grouper (`src/report/group-issues.ts`) collapses issues that share check + severity + normalized summary (paths/URLs/viewports/digits stripped), so "description ausente" on 10 pages reads as one row with an affected-pages note instead of 10 near-identical rows. Applied to the HTML Top issues card, the CLI Top issues list, and the PR comment. Issue IDs and baselines are untouched.
+* **Module selection: `--only`/`--skip`/`--why`.** `parity run` now groups its ~30 checks into 8 named modules (`e2e`, `seo`, `visual`, `vitals`, `cache`, `console`, `html`, `network`). Scope a run with `--only e2e,seo` (or single-check granularity via `check:<name>`), subtract with `--skip`; unselected flows/sitemap-crawl/visual-diff passes are skipped entirely rather than run-and-discarded, making scoped runs genuinely lighter and faster. `--why <text>` records the scoping rationale in `report.json`. No selection at all = full back-compat (every check runs, exactly like before). `parity list modules [--json]` discovers the taxonomy programmatically. A TTY with no explicit selection shows an interactive prompt; non-TTY just gets a one-line heads-up and proceeds with everything.
+* **Adaptive, per-module scoring.** The verdict score now reflects only the modules that ran: `--only e2e` scores purely on e2e; `--only e2e,seo` blends both, weighted by pages actually analyzed per module. `report.json` carries the full breakdown (`moduleVerdicts`), `verdict.modulesRun` lists contributing modules, and score-trend comparisons only match runs that scored the *same* module set.
+* **Cart interactions overhauled.** Multi-item add + validation (`add-second-item`/`validate-multi-item`), cart persistence across reload (`verify-cart-persistence`), direct quantity input (`set-qty-input`), configurable coupon codes (`rc.coupon`) with an opt-in valid-coupon assertion (`apply-valid-coupon`), and a VTEX-only informational probe (`seller-code-null`, issue: typing the literal string `"null"` into a seller-code field is accepted by VTEX and resolves the seller itself) — always `ok`/`skipped`, never fails the run.
+* **Interactive PLP pagination.** The `plp` flow now drives pagination like a user would — clicking a next-page link, clicking "load more", or scrolling — instead of only fetch-probing `?page=N`. Catches load-more/infinite-scroll sites the old fetch-only check couldn't, while keeping the fetch probe as a fallback for classic paginated PLPs.
+* **New checks:** `pdp-breadcrumbs` (breadcrumb trail or JSON-LD `BreadcrumbList` still renders), `plp-sorting` (a sort query param actually reorders products), `spa-navigation-flow` (F5 vs client-side `<Link>` navigation — catches CMS sections/site-globals silently dropping on SPA nav, a real bug class from a production Fresh→TanStack migration post-mortem), `serverfn-hover-flood` (hovering product cards shouldn't flood the worker with more than a configurable budget of server-fn/preload requests).
+* **Selector discovery v2.** Discovery now sees up to 3 labeled HTML sources (home + PLP + PDP) instead of home alone, so PDP-only and PLP-only selector keys are grounded in the page they actually live on. The model also flags low-confidence guesses (`low_confidence_keys`). A live-validation pass (`page.locator(sel).count() > 0`) runs before a selector is trusted or promoted to the learned-selectors library — a selector that validates AND wasn't flagged low-confidence is promoted directly as `origin: "verified"`; one that fails validation is dropped from the run entirely rather than silently misfiring later. `parity learned validate --url <url>` exposes this as a standalone diagnostic. Selector cache gained TTL + structural-fingerprint invalidation (previously cached forever until manually refreshed).
+* **`parity extract`** — a new single-site command (no prod×cand comparison) that captures AI-ready structured data about a site's UI components (header, footer, nav, shelves, hero, minicart, …) for migrations where there's no source code to read: HTML, computed styles, cropped screenshots, asset/link inventories. Heuristic component auto-detection (semantic markup + Deco `data-section` convention + geometry, with an optional LLM relabeling pass) plus pluggable exporters (Markdown for pasting into an agent's context, JSON manifest).
+
+### Changed
+
+* **`computeVerdict` is now a single shared module** (`src/engine/verdict.ts`) — `run.ts`, `vitals.ts`, and `cache.ts` previously carried three drifting copies.
+* **`src/engine/flows.ts` (3.9k lines) split into `src/engine/flows/`** (one file per flow + shared helpers) — no behavior change, just a maintainability split ahead of the M2 feature work above.
+* Learned-selectors gained a lifecycle: entries now carry `origin: "verified" | "llm-guess"`, staleness decay based on `lastValidated` (previously stored but never read), and verified-first ranking.
+
+### Fixed
+
+* **`lazy-section-presence` false positive for Fresh→TanStack/Vite chunk naming** (issue #118) — Fresh names a lazy-loaded section chunk `render`, Vite names the same chunk `render.ts`; the check now strips bundler extensions before comparing, so it stops reporting a false "missing section" on every page with lazy sections.
+
 ## [0.11.16](https://github.com/decocms/parity/compare/v0.11.15...v0.11.16) (2026-06-17)
 
 ### Changed