feat: compare databases, mirror volumes and seed profiles

[machado144]

Summary

• compare: row counts, size, delta and column drift per table between any two connections. Read-only on both sides; tables match across engines ignoring case.
• mirror: seeds a target so its volumes follow a source (--scale, --max-rows, --tables, topup or reset). --dry-run shows the plan and sample rows. Refuses when source and target are the same database.
• Seed profiles: value rules applied by column pattern or per column. Actions are template ({{auto}}, {{seq}}, {{run}}, any generator), faker, value, oneOf and setNull. Profiles are saved to ~/.config/seedstorm/profiles.yaml and used via --profile on seed/gaps/generate/mirror, in the TUI, and in the web UI.
• Resilient inserts (internal/seeder): tables fill in chunks and rejected rows are regenerated. A table that can't make progress is reported as partial/failed and the run continues.
• Large volumes with flat memory: seed, gaps and mirror generate and write 20k rows at a time (one shared seeder.Seed engine for CLI, TUI and web); Postgres chunks go through COPY; key state is read once per table with sampled PK pools and Bloom-filter key sets.
  • 7.2M-row mirror: 3m23s / 1.48GB peak → 2m27s / 147MB
  • seed of 600k rows on Postgres: 41.9s / 585MB → 7.3s / 97MB
• Streaming generate/export (internal/dataio): files are written and read a chunk at a time, same YAML/JSON layout as before. 300k rows: generate 1.65GB → 99MB, export 2.14GB → 80MB. CSV added to generate, YAML/JSON input to export.
• Bounded serve: job output to the browser caps at 20MB (dry-run SQL notes what it left out; generate/export refuse and point at --out), job requests are size-limited, and idle session connections close after 2 minutes.
• Faster MySQL inserts: default batch size 100 → 1000 (200k rows: 36s → 9s); batches split to stay under 65,535 placeholders and ~1MB.
• Postgres sequences behind SERIAL/IDENTITY columns are advanced after seed, gaps and mirror, so the application's own inserts don't collide with seeded ids.
• Web: new Compare and Profiles pages, a profile select in the workspace, and a nav drawer on narrow screens.
• Re-seeding a populated DB now appends: ids, UNIQUE sequences and composite keys continue past existing rows.

Fixes found while testing against Keycloak and the DB version matrix:

• *_time/*_date columns stored as integers got time strings
• string PKs ignored varchar(n)
• MySQL clone-schema lost index prefix lengths
• MySQL 5.7 rejected a cloned nullable TIMESTAMP
• multi-column UNIQUE constraints (Keycloak's UNIQUE (realm_id, username)) were ignored; tuples are now kept distinct and rows that can't be made distinct are dropped with a warning
• explicit profile table/column names didn't match MySQL's upper-case identifiers; value: true failed on MySQL TINYINT(1)
• --counts estimate reported unknown or stale-zero counts on never-analyzed tables (pg13); those tables are now counted exactly
• serve re-downloaded ~870KB of scripts on every page; static assets now revalidate with ETags
• SQL from generate --format sql, export --format sql and dry runs had $1/? placeholders and no values; it now writes literals escaped per engine and loads as-is

CI integration timeout is raised to 900s; the suite now takes about 5 minutes.

seedstorm compare --source-dsn "$PROD" --target-dsn "$STAGE" --only-diff
seedstorm mirror  --source-dsn "$PROD" --target-dsn "$STAGE" --scale 5 --max-rows 2000000 --profile loadtest --dry-run

name: loadtest
rules:
  - column: "*email*"
    template: "lt+{{seq}}.{{run}}@example.test"
tables:
  users:
    columns:
      role: { value: guest }

Test plan

• Unit tests (go test -race ./...) for rules, compare/plan, seeder retry decisions (scripted driver), profiles store, faker existing-row handling, TUI mirror model, and web APIs
• Binary-driven integration evals on Postgres + MySQL: re-seed, compare/mirror end to end, seeder rejections, Keycloak 87-table schema (pg, mysql, mysql→pg), web compare/mirror/profiles
• Full integration suite green (282 subtests); new evals also green on pg13 + mysql5.7 and pg17 + mysql8.4
• Real replication with every rule kind: Keycloak on pg15 → fresh postgres:17 container, and mysql8.0 → fresh mysql:8.4 container; 1× and 2× exact on all 87 tables, 0 custom-field violations in the target; generate --profile checked on both
• Compare and Profiles pages driven in Playwright; zero horizontal overflow at 320–1920px
• Evals: seed/gaps --fill of 300k rows under 150MB peak (old code: 299MB) and an 80-column table at the default batch size (fails on MySQL without placeholder splitting)
• Evals: generate/export of 300k rows under 150MB; exported and generated SQL loads on both engines with quotes, backslashes, newlines, unicode and NULL intact (drop the MySQL backslash escaping and it fails)
• Web generate → Export this → export SQL, dry runs and the 20MB refusal driven in Playwright
• 7.2M-row mirror benchmark (accounts → orders → order_items): no orphans, unique emails, CHECKs hold, sequences advanced
• golangci-lint, gofumpt, gauntlet (structlint, dupehound on changed lines)

[github-actions]

✅ PR title follows the required format

Current title: feat: compare databases, mirror volumes and seed profiles