feat(applications): visual schema.graphql editor

[dawsontoth]

Closes #1417.

A real GUI for editing Harper schema.graphql files. Opening a .graphql file now shows a visual, table-oriented editor by default, with an "Edit as text" escape hatch to Monaco — the same preview/source toggle pattern as the SVG editor. This subsumes the hidden File → New Table modal (which only ever appended a table).

Stacked on # — base branch is feat/edit-svg-as-text (it reuses that branch's view-toggle plumbing). Please merge that first, or review this against its base.

What it does

• Visual-first: .graphql files open in the table editor; toggle to raw text anytime.
• Add / edit / remove tables and fields (not just add).
• Full directive coverage, driven by the existing harper-language metadata: @table (incl. database, expiration, audit, eviction, scanInterval, replicate, randomAccessFields), @export (name/rest/mqtt), @sealed, and every field directive — @primaryKey, @indexed (incl. vector/HNSW params), @computed, @relationship, @allow, timestamps, @expiresAt, @enumerable. (Reconciled the metadata mirror to add the missing randomAccessFields.)
• Collapsible table cards — existing tables open collapsed with a summary (name · N fields · REST); newly-added tables open expanded.
• GraphQL-name validation highlights invalid table/field names (red field + hint + header flag).

Safety / fidelity (issue's #1 ask)

The core is a small, dependency-free, well-tested parse/serialize library (lib/schema/) that round-trips a schema losslessly:

• Unedited tables and anything the GUI doesn't model — comments, scalars/enums, non-@table types, unknown directives/args — are preserved verbatim. Only edited/new tables are regenerated canonically.
• Genuinely malformed SDL falls back to the text editor rather than risk mangling it.
• An unnamed, in-progress table isn't written to the file until it has a name (no type @table orphan; doesn't even mark the file dirty).

Stability (issue's #3 ask)

Generated output mimics the file's indentation (tabs vs N spaces) and orders tables alphabetically; regeneration is idempotent.

Testing

• lib/schema/ unit tests: round-trip fidelity, unknown-directive/arg + comment preservation, tabs-vs-spaces, alphabetical ordering, list/non-null, the mutation→serialize pipeline, skip-unnamed-table, and GraphQL-name validation.
• TableCard component tests (jsdom): render/edit/toggle/collapse/read-only.
• Full suite green (1187 passing), typecheck + lint clean.
• Browser-verified against real cluster data: parse a multi-table schema, add/edit/remove, toggle directives, collapse, invalid-name highlight, unnamed-table handling, and light + dark mode.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces a visual-first schema editor (SchemaEditorView) for Harper schema.graphql files, replacing the previous modal-based table creation form. It includes parsing and serialization logic to seamlessly toggle between the visual editor and a raw text editor while preserving comments, formatting, and unmodeled schema definitions. Feedback on the changes highlights a few key improvements: using stable keys instead of array indices for the dynamic list of fields to prevent UI state leakage, only sorting tables alphabetically during serialization when edits are present to avoid marking files as dirty immediately upon opening, and defensively skipping empty-named fields during serialization to prevent invalid GraphQL syntax.

[kriszyp]

Reviewed via Claude review-queue. Genuinely well-architected feature — the parse → segment model → serialize approach that keeps everything unmodeled as byte-for-byte raw text is the right design, and the visual/text stale-overwrite race I was watching for is correctly handled (bufferRef captures live edits before the reset effect runs).

Two real data-fidelity bugs in GraphQL description ("""...""") handling, both empirically confirmed by running the actual PR-head parse/serialize round trip, and neither covered by the test suite:

Blocker: multi-line description gets corrupted with duplicate indentation the first time its table is edited. parseFields captures a field's """...""" description with its original embedded indentation, then generateTable blindly re-prepends the indent to every line on regeneration. Confirmed: editing any field on a table with a multi-line description doubles the indent on every interior line, and doing it twice doubles it again.

Blocker: a table's leading """...""" description doesn't travel with it on alphabetical sort. splitAttachedLeading only recognizes # comment lines as "attached" trivia that moves with a table when tables get re-sorted — a preceding docstring isn't matched, so it stays anchored at the old position and ends up describing whatever table sorts into that slot instead. Confirmed with a two-table repro.

Since lossless round-tripping is the PR's own stated #1 ask, I'd like these fixed (plus regression tests covering multi-line descriptions at both table and field level, through an edit + sort) before merging. The pre-existing silent save-error-surfacing gap is worth a follow-up ticket but not a blocker for this PR.

[dawsontoth]

Thanks for the thorough round-trip testing — both description bugs are fixed in a0a36de, with regression tests through edit + sort:

• Multi-line field description indent doubling: generateTable now strips any indentation a comment/description line already carries before applying the canonical indent (blank lines stay blank), so regeneration is idempotent — editing twice no longer compounds it.
• Table leading description not traveling on sort: splitAttachedLeading now recognizes """…""" blocks (not just # lines) as attached trivia, so a docstring moves with its table when tables are reordered, while unrelated content above a blank line stays put.

New tests: verbatim round-trip of single-/multi-line descriptions at table and field level (parseSchema.test.ts), field-description edit idempotence, and a table-description alphabetical-sort repro (serializeSchema.test.ts). Full suite green (1192 passing), typecheck + lint clean.

Filing the pre-existing silent save-error-surfacing gap as a separate follow-up as suggested.

🤖 Addressed by Claude Code

[kriszyp]

Reviewed via Claude review-queue — both prior blockers (double-indent on multi-line descriptions, docstring not traveling with table on resort) fixed and test-verified (36/36 vitest incl. 2 new regression tests). No new bugs. LGTM.