fix(tools): server-side case-insensitive normalize of gridItems[].args keys

[romer8]

Summary

Description-tightening (PR #13) didn't stop gemini-flash from emitting `args = {"river_id": ...}` when the plugin declared `river_ID`. Per the escalation path agreed in /ce-debug Option B, this PR adds server-side case-insensitive normalization: fetch the source's declared `arg_names` from `TETHYSDASH_BASE_URL` and rewrite mismatched gridItem args keys to canonical case before persisting.

Root cause (confirmed in two LLM runs)

User prompt: "embed the GeoGLOWS Forecast Plot plugin with River ID = ${feature.comid}". The plugin declares `arg_names: ["river_ID"]`. The LLM emitted `args_string: "{\"river_id\": ...}"` (lowercase). Plugin's `run()` looked up `river_ID` → undefined → no data fetched. Edit Visualization "River ID" form input bound to key `river_ID` → empty.

PR #13 added "case-sensitive" + the `river_ID` counterexample to the description. gemini-flash still emitted lowercase. Per workspace memory:

`feedback_proactive_over_reactive_llm_routing.md`: "Pattern-matching against open-vocabulary model output is a losing fight."

Description tightening worked for the cols bug (Bug B from prior /ce-debug; LLM correctly emitted `w=100` after PR #13). It didn't work for case-sensitivity because LLM snake-case normalization is a much stronger prior.

Fix

Two new helpers in `mcp_server.py` near `configure_popup_modal_layer`:

`_fetch_plugin_arg_names(source)`

Hits `${TETHYSDASH_BASE_URL}/visualizations/list/` (same endpoint as `list_intake_plugins`). Returns:

• `list[str]` when source found with declared args
• `[]` when source found but has no declared args (Default registry types like Map/Text whose args are open-shape)
• `None` when source not found OR fetch failed

`_normalize_args_case(args, arg_names)`

Case-fold matches LLM-emitted args keys against `arg_names`. Rewrites mismatched keys to canonical case. Soft rules:

• `arg_names` is `None` / empty → pass through unchanged
• Exact match → preserved
• Case-fold match → rewrite (e.g., `river_id` → `river_ID`)
• No match for a key → keep as-is (cheap path; runtime tile error boundary handles bad arg downstream)
• Two LLM keys collide on case-fold → return `None` → structured envelope error

Wiring

In `configure_popup_modal_layer`'s gridItem normalization loop: for each gridItem, fetch its source's arg_names and normalize args in-place before `json.dumps` into `args_string`.

Soft-fails to pass-through when `TETHYSDASH_BASE_URL` is unset or fetch fails — better to ship the LLM's verbatim args than to reject the whole flow when the registry is unreachable.

Bonus: tighter h-hint per user feedback

User noted "the height can be reduced a bit" — LLM picked `h=40` which was visually too tall. Refined description's h guidance:

• `h~25-30` for a single time-series plot (common case)
• `h~35-40` for a tall plot or one with thick legends
• `h~15-20` for a card or short text block
• avoid `h>50` unless the popup will scroll

Tests

7 new in `TestPopupConfigArgsCaseNormalization`:

• `test_lowercase_key_rewritten_to_canonical_case` — motivating real-user case
• `test_exact_match_preserved`
• `test_unknown_source_args_pass_through`
• `test_default_registry_source_no_declared_args`
• `test_fetch_failure_soft_fails_to_pass_through`
• `test_case_fold_collision_rejected_with_structured_error`
• `test_normalization_runs_per_gridItem_independently`

Full suite: 944 passed (937 baseline post-#13 + 7 new). No regression — existing tests pass because `TETHYSDASH_BASE_URL` is unset in the test env so `_fetch_plugin_arg_names` returns `None` and the helper passes through.

Trade-offs

• Network call per `configure_popup_modal_layer` invocation (per gridItem source). No caching in v1; can add per-request memoization or TTL cache if observed latency justifies.
• Soft network-failure behavior — when fetch fails, args pass through verbatim. The LLM's wrong-case key still gets persisted (current pre-fix behavior). Better than blocking the flow when the backend is briefly unreachable.
• Doesn't generalize to the `attribute_variables` arg on `add_*_layer` tools (different bug class but same pattern). If users hit the same case-mismatch issue there, the helper is reusable.

Manual smoke after merge

Restart MCP server. Re-run turn-2 prompt. Expected:

• args_string contains `"river_ID"` (capital ID) regardless of what the LLM emitted
• Popup opens at runtime with data fetched
• Edit Visualization modal "River ID" field populated with `${feature.comid}`
• gridItem h closer to 25-30 (per the tighter description hint)