RodexDB v0.5.0

Added

• Capacity modes (platform-wide): PERFORMANCE = all tables on-demand
  (pay-per-request, no throttling, budgets become guardrails: writes
  100 000 units/min, reads 400 000/min) · NORMAL = provisioned 5/5 ($0 free
  tier, budgets 800 write-units / 800 reads per app/min — the free-tier pool made honest).
• Switch via dashboard (AppsPage strip + confirm modal, SWITCHING… state)
  · REST GET/POST /v1/admin/capacity · MCP get_platform_capacity
  • set_platform_capacity (confirmation-gated) — 26 tools.
• Item cap 20 KB → 400 KB (DynamoDB's hard limit) in BOTH modes — reads
  were never size-gated; writes now accept any size; 20 KB becomes a cost
  recommendation. Batch byte cap 400 KB total.
• docs/capacity.md (matrix + cost ladder + AWS switching facts).

Changed

• Budgets in NORMAL were raised (120 → 800 write-units/min per app; 240 → 800 reads; 600 → 2 000 total;
  1 000 → 2 400 platform — the real 25 WCU+25 RCU/s shared pool with margin) — 429s
  now only at genuinely absurd rates.
• New tables created while PERFORMANCE are on-demand directly.

Tests

• 177 gateway + 14 package (capacity endpoint suite, PERF guardrails,
  400 KB round-trip, TEST profile for exact end-to-end bursts).

Fixed (from the write-safety round)

• CORS: DELETE now allowed from the console — app soft-delete and MCP
  key delete were blocked by preflight since launch (day-one bug).
• delete_table pacing — drains big-row tables with ≤20-item chunks,
  1 s gaps and 429-retry backoff; cleanup no longer throttles mid-drain.

Changed (WCU-honest writes — the universal precaution, research-validated)

• Write budget is now size-honest: 120 write-units/min where every
  row costs max(1, ceil(bytes/1024)) — DynamoDB's exact rounding rule.
  Rows ≤ 1 KB = 1 unit (unchanged); an 18 KB row = 18 units; 429s name the
  budget. Matches the OpenAI TPM / DynamoDB WCU capacity-unit precedent.
• batch/put byte cap: total serialized bytes ≤ 20 KB per call, checked
  before any write → 413. A WCU-burst batch is structurally impossible.
• all_ok flag on batch responses (Elasticsearch errors precedent):
  a 200 with failed items is now impossible to miss. Per-item errors still
  carry ok:false + reason; retry guidance documented everywhere.
• bytes echoed on every item (put/get/query/update/batch) — consumers
  see the exact number the budget charges on.

Tests

• 183 → 188 (all_ok matrix, batch byte cap, WCU-unit ceiling proof,
  bytes echo, CORS preflight regression).