fix: chart time-axis rendering, auto-detection, and MCP ergonomic fixes (#39)

* fix(mcp): disable ANSI escape codes on stderr tracing layer

MCP servers communicate over stdio; their stderr goes to a client log
pane (e.g. Claude Code output panel) that doesn't render ANSI codes.
On Windows this produced raw escape sequences like [2m...[0m. Force
plain text on both the daemon and MCP stderr layers.

Also adds a pre-1.0 stability note to the README.

* fix(mcp): auto-create parent dirs for chart/export output paths

validate_output_path previously rejected paths whose parent directory
didn't exist. Now it creates the directory tree (create_dir_all) before
canonicalizing, so users don't need to pre-create output directories.

Also improves chart tool description to guide LLMs on:
- Long-format data requirement (one numeric y, use series for grouping)
- x_as_category=true requirement for DATE/TIMESTAMP x-axes
- UNION ALL reshaping for wide-format data

* fix(chart): auto-detect categorical x columns for line/scatter charts

Line and scatter charts now peek at the first row's x value — if it's
not numeric (DATE, TIMESTAMP, TEXT, etc.), categorical mode activates
automatically. This eliminates the need to manually set
x_as_category=true for date-based time series.

The explicit x_as_category parameter still works as an override.

* fix(chart): auto-shorten TIMESTAMPTZ labels and thin dense tick marks

Categorical axis labels now go through a shortening pass:
1. Strip shared timezone offset (+00:00) when all labels end with it
2. Auto-thin: blank intermediate labels when they'd overlap based on
   estimated character width vs chart pixel width

This prevents label collision on charts with 12+ TIMESTAMPTZ ticks at
800px width. First and last labels are always preserved.

* feat(chart): proportional time axis for DATE/TIMESTAMP/TIMESTAMPTZ + fix tick thinning

Line and scatter charts now auto-detect DATE / TIMESTAMP / TIMESTAMPTZ
x columns and render with a proportional time axis: real wall-clock
gaps between data points are reflected on the chart's x-axis instead
of being flattened to even spacing. Tick labels are formatted via
chrono to match the input kind (%Y-%m-%d, %Y-%m-%d %H:%M:%S, or
%Y-%m-%d %H:%M:%S%:z), with the originating offset captured from the
first row preserved for TIMESTAMPTZ. Pass x_as_category: true to opt
out and force the previous categorical layout; TEXT x columns
continue to render categorically.

Also fixes a tick-label thinning regression in the categorical path.
The old shorten_labels blanked individual labels at non-step indices,
but plotters picks its own tick positions on the float axis and
rounds them to integer indices, so almost none of the chosen ticks
landed on a kept index - a 90-point hourly TIMESTAMP series at any
chart width rendered with only ONE visible x-axis label. The chart
layer now computes a target tick count from chart width and label
size, then passes it to plotters via .x_labels(N); every drawn tick
carries a real label.

Refactors:
- shorten_labels split into strip_shared_tz_suffix (cosmetic +HH:MM
  compaction) and auto_tick_count (no-overlap budget for plotters).
- group_series x_as_category: bool -> x_mode: XMode enum (Numeric /
  Categorical / Temporal(TemporalKind)).
- New parse_temporal, format_temporal_tick, detect_line_x_mode,
  tick_count_for_label_width helpers.

Tests: 16 new unit tests for the helpers (parse, format, detect,
tick math, NaN safety, offset capture); 1 new integration test
(line_chart_long_timestamp_series_renders_multiple_visible_labels)
asserts >=3 visible date labels in the rendered SVG, locking in the
regression fix. Total: 26 unit, 37 integration, all green; clippy
--all-targets -D warnings clean.

Docs: chart tool description in server.rs and LLM-facing readme.rs
updated to document the new auto-detect behavior; CHANGELOG.md
[Unreleased] gains two ### Fixed entries.

* chore: re-enable release-please to release v0.2.0

* chore: fix formatting

---------

Co-authored-by: Stefan Steiner <ssteiner@salesforce.com>

Author: StefanSteiner

.github/workflows/release-please.yml

@@ -28,9 +28,7 @@ permissions:
 
 jobs:
   release-please:
-    # Temporarily disabled until post-daemon/two-db-model work is complete.
-    # Re-enable by removing the `false &&` prefix below.
-    if: false && github.repository == 'tableau/hyper-api-rust'
+    if: github.repository == 'tableau/hyper-api-rust'
     runs-on: ubuntu-latest
     steps:
       - uses: googleapis/release-please-action@v5

README.md

@@ -10,11 +10,11 @@ A **pure-Rust** implementation of the Hyper database API, using the PostgreSQL
 wire protocol with Hyper-specific extensions. Create, read, and manipulate Hyper
 database files (`.hyper`) without any C library dependencies.
 
-> **Project Status — 0.1.x, AI-Engineered**
+> **Project Status — 0.2.x, AI-Engineered**
 >
 > This crate was vibe-engineered with heavy use of AI coding assistants. The
-> **0.1.x** line may still undergo large breaking changes; the public API
-> won't settle until the 0.2.0 release.
+> **0.2.x** line may still undergo large breaking changes; the public API
+> won't settle until the 1.0.0 release.
 >
 > Contributors and reviewers should, at a minimum, run an **AI code reviewer**
 > over any changes, following the conventions, layering rules, and patterns

hyperdb-mcp/CHANGELOG.md

@@ -143,6 +143,35 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 
 ### Fixed
 
+- **Chart x-axis tick label thinning.** Long categorical line/scatter
+  charts (e.g. a 90-point hourly TIMESTAMP series) used to render with
+  only ONE visible x-axis label. The old logic blanked individual
+  labels at non-step indices, but `plotters` picks its own tick
+  *positions* on the float axis and rounds them to integer indices —
+  so almost none of the chosen ticks landed on a kept index, and the
+  formatter returned empty strings for the rest. The chart layer now
+  computes a target tick count from chart width and label sizes and
+  passes that count to `plotters` via `.x_labels(N)`; every drawn
+  tick carries its real label. Same `+00:00` suffix stripping for
+  shared TIMESTAMPTZ offsets is preserved (now isolated in
+  `strip_shared_tz_suffix`).
+- **Line / scatter charts over `DATE`, `TIMESTAMP`, and `TIMESTAMPTZ`
+  columns now use a proportional time axis** instead of the previous
+  categorical-with-evenly-spaced-ticks behavior. Real-world time gaps
+  between data points are now reflected in the chart's x-axis: a
+  series at `2026-05-01 08:00`, `2026-05-01 12:30`, `2026-05-02 06:15`
+  shows the 4.5-hour and 17.75-hour gaps proportionally instead of
+  flattening every interval to the same width. Tick labels are
+  formatted via `chrono` in a form that matches the input kind:
+  `%Y-%m-%d` for DATE, `%Y-%m-%d %H:%M:%S` for TIMESTAMP, and
+  `%Y-%m-%d %H:%M:%S%:z` for TIMESTAMPTZ (the offset captured from the
+  first row, so a uniformly-`+05:30` series reports IST throughout).
+  Set `x_as_category: true` to opt out and force the previous
+  categorical layout (e.g. for charts where evenly-spaced bins are
+  more readable than proportional gaps). TEXT x columns continue to
+  render categorically as before. Bar charts are unaffected — they
+  remain categorical regardless of x type, which matches reader
+  expectations for grouped data.
 - **Watcher recovery after hyperd restart.** The watcher's connection
   pool now auto-rebuilds when a per-file ingest hits a connection-lost
   error (typically after the daemon restarts hyperd). Each ingest gets

hyperdb-mcp/README.md

@@ -1,5 +1,7 @@
 # hyperdb-mcp
 
+> **Note:** This crate was vibe-engineered with heavy use of AI coding assistants. The 0.1.x line may still undergo large breaking changes; the public API won't settle until the 1.0.0 release.
+
 An MCP (Model Context Protocol) server that turns the Hyper columnar database into an instant SQL analytics engine. Data flows in from other MCP plugins or files, lands in Hyper automatically, and becomes queryable with SQL — no setup, no schema files, no database management.
 
 Built on the pure-Rust [`hyperdb-api`](../hyperdb-api/) crate for maximum performance: 22M+ rows/sec inserts, 18M+ rows/sec queries, constant memory for billion-row results.

hyperdb-mcp/src/attach.rs

@@ -727,6 +727,14 @@ pub fn validate_output_path(path: &str, kind: &str) -> Result<PathBuf, McpError>
             format!("{kind} path '{path}' has no file-name component"),
         )
     })?;
+    if !parent.exists() {
+        std::fs::create_dir_all(parent).map_err(|e| {
+            McpError::new(
+                ErrorCode::InternalError,
+                format!("Failed to create parent directory for {kind} path '{path}': {e}"),
+            )
+        })?;
+    }
     let canonical_parent = std::fs::canonicalize(parent).map_err(|e| {
         McpError::new(
             ErrorCode::FileNotFound,
@@ -924,14 +932,19 @@ mod tests {
     }
 
     #[test]
-    fn validate_output_path_rejects_missing_parent() {
-        // Build a platform-portable absolute path with a missing parent.
-        // Hardcoded "/definitely/..." paths are not absolute on Windows.
-        let missing_parent = std::env::temp_dir()
-            .join("hyper_mcp_validate_output_missing_parent_99999")
-            .join("out.csv");
-        let err = validate_output_path(missing_parent.to_str().unwrap(), "export").unwrap_err();
-        assert_eq!(err.code, ErrorCode::FileNotFound);
+    fn validate_output_path_creates_missing_parent() {
+        let parent = std::env::temp_dir().join("hyper_mcp_validate_output_missing_parent_99999");
+        // Clean up from any prior run.
+        let _ = std::fs::remove_dir_all(&parent);
+        assert!(!parent.exists());
+
+        let target = parent.join("out.csv");
+        let canonical =
+            validate_output_path(target.to_str().unwrap(), "export").expect("should create parent");
+        assert!(canonical.is_absolute());
+        assert!(parent.exists(), "parent directory should have been created");
+        // Clean up.
+        let _ = std::fs::remove_dir_all(&parent);
     }
 
     #[test]