Remove the second ordering engine
The demo initially reimplemented grouping, dependency traversal, priority propagation and selection order. Its first 15 review rounds recorded 31 findings, including regressions introduced while repairing earlier findings.
Once the shared derive package shipped, the demo was reduced to a projection between its document and the library’s inputs and outputs. The duplicated graph logic was deleted.
At the intermediate rework (44619a2 → 2977a56), demo/src/order.ts fell from 997 to 801 total lines and order.test.ts from 1,042 to 504, including comments and blank lines. Later review changed both files again. Across the complete PR, 18 published review submissions raised 39 finding reports. The final disposition retained a source-state follow-up and carried three cycle findings into package issue #43.
Find a deadlock that individual issue checks missed
Three issues could wait forever when a blocking cycle passed through a together unit, while the cycle list and diagnostics remained empty and the pre-write guard returned false.
PR #46 made the schedulable unit the graph vertex, updated both cycle reporting and the write guard, and clarified the specification. Internal advisory edges and singleton self-loops remain distinct.
The PR records the three-issue reproduction, four deliberate reversions each caught by the relevant test suite, and the demo’s known-gap test failing as intended when the package caught up. Current reader and derive tests preserve the group-contraction and self-loop cases.
A write result says whether the requested edit landed
A null splice result conflated no block with an existing uneditable block, and successfully parsing the output did not establish that the requested relationship was written. Provenance relationships could also be set but not cleared.
PR #55 introduced distinct result types and a read-back check of the requested edit. PR #65 made omitted/set/clear ownership uniform across the four writable fields, keeping unrelated text unchanged.
Reintroducing the historical indentation bug changed false success into not-written. The PR records 25 tests failing when that check is removed with the old defect. The later compatibility check compared 23 unchanged CLI cases and four newly permitted cases; a review correction added stderr to stdout and exit-code comparison.
Verify what consumers can actually install
Workspace checks passed while published reader/writer packages imported an export missing from core 0.1.0. A static resolver audit then accumulated findings while trying to model package-manager and runtime behavior.
The release path now installs the exact intended versions in a fresh consumer, inspects the entire resolved dependency tree, imports every published package and exercises the command.
PR #51 records a seven-package release validation and one package needing six propagation attempts while another was immediately available. That is one historical registry observation, not a reliability benchmark. Current tests cover every publishable package, stale nested copies, absent packages and CLI failure. The inventory is now eight published packages.
A retry against fresh data is one operation
Hosts had to compose a refresh and retry across an asynchronous gap. A refresh failure or a second action during that gap could make the retry use stale state or race user intent.
PR #110 moved retry-on-latest into the store’s existing queue and reserved the edit before the read. A failed read restores the original record and sends nothing.
Current tests directly cover a failed refresh, discard and second-press attempts during a held read, a refreshed document that invalidates the edit, and ordering behind an in-flight write. The public PR records these cases and the existing pending state being reused rather than adding a new queue.
Test the runtime promise by calling the packages
All eight published packages declared Node 18 support, but their build assumed ES2023 and the compatibility check only imported package entry points. Unsupported built-ins inside a function could remain invisible.
PR #233 aligned the compiler with ES2022 and changed the existing smoke check to call a public function from every package with a declared runtime floor.
Lowering the compiler library found findLastIndex in an editor test. The PR records 32 smoke tests and three failures when the new call requirement was disabled. Current tests require a probe for every published package and fail a probe that throws. The public record explicitly says this exercises selected paths, not every possible runtime behavior.
