fix(mcp): reduce LLM-facing tool friction, token cost, and a scratch-table leak

[StefanSteiner]

Summary

Makes the HyperDB MCP more reliable and cheaper for an LLM to drive. Came out of a systematic pass exercising all 33 tools live and tracing the friction points back to source. Two axes: fewer LLM mistakes, and lower always-in-context token cost.

Bugs fixed

• Scratch-table leak: query_data / query_file dropped their _tmp_* staging table only after the query ran, so a failed query orphaned it — and it then showed up in describe as a mystery table. The DROP now runs whether the query succeeds or fails.
• Substring-collision in SQL rewrite: the table-name substitution was a naive sql.replace(alias, temp) that corrupted a column like metadata when the alias was data. Replaced with a whole-word replace_identifier helper (unit-tested). Known limitation documented: it is identifier-boundary aware but not SQL-aware, so a string literal exactly equal to the alias is still rewritten — rare, since the alias is a table name.
• describe clutter: table listings now hide transient scratch tables (_tmp_*, __hyperdb_merge_*), not just _hyperdb_* internals.

Ergonomics

• query_file default table is now data (matching query_data and the documented FROM data example, which previously failed because the default was the unpredictable file stem).
• kind is now optional on attach_database and copy_query's temp_attach, defaulting to local_file — no more required ceremony field with only one legal value.
• Better unknown-table error: a 42P01 now suggests running describe and names the data alias, instead of the generic "check SQL syntax."

Token cost

• Trimmed the essay-length tool descriptions (load_file, export, chart, load_files, set_table_metadata, kv_set + 3 updated): ~15.8k → ~11.7k chars (~26%) of always-in-context tool-catalog text. The format-selection and edge-case depth moved into get_readme (consolidated format-preference block), which stays under its 24k budget.

Type-name casing normalization (TEXT vs text) was deliberately deferred — it originates from two engine code paths asserted on by many tests, and type names are case-insensitive so it causes no actual LLM mistakes (high-risk / low-value).

Test plan

• cargo test -p hyperdb-mcp --all-targets against pinned .hyperd/current — 669 tests, 0 failures (incl. new replace_identifier and is_internal_table unit tests)
• cargo clippy --workspace --all-targets --all-features -- -D warnings — clean
• cargo fmt --all — no diff
• npx markdownlint-cli2 on the touched CHANGELOG — 0 issues
• Independent code review — approved merge-ready
• All 33 tools exercised live against a running server