Formatting on save

Build the release formatter from the repository root:

cargo build --release -p zincite-fmt --bin zincite-fmt

Configure your editor’s external formatter with the absolute path to target/release/zincite-fmt and these arguments:

--stdin-filepath /absolute/path/to/the/edited/model.mzn

Send the current in-memory buffer as UTF-8 on stdin. Use the actual destination path, even for a new file that has not been saved yet. A .dzn path selects data syntax; other paths select model syntax. The path also selects the EditorConfig lookup starting directory and appears in diagnostics. No input argument is needed: the command reads stdin by default. Plain stdin without --stdin-filepath uses model syntax and default settings without filesystem configuration lookup.

The default style uses four spaces, LF, a final newline and a 120-column limit. EditorConfig supplies indentation, tab width, editable line endings, width, final-newline and trailing-whitespace settings, and UTF-8/BOM selection. Supported command-line style overrides take precedence; see --help for their values and the formatting style gallery for examples. Formatting may change between versions before a stable release.

Editor recipe

1. Take the edited buffer and its destination path at the start of the save.
2. Start one fresh formatter process, passing the path as one argument. Send the complete buffer on stdin and capture stdout and stderr separately.
3. Wait for process completion and complete stdout capture. On exit status 0, replace the buffer with the captured stdout, then save it. An already formatted buffer produces the same bytes. Successful output may be empty for empty input.
4. On nonzero status, timeout, launch failure or incomplete output capture, keep the original buffer and show stderr or the process failure through the editor’s usual diagnostic UI. Follow the editor’s existing policy for saving that unchanged buffer.

Do not stream stdout into the buffer, merge stderr into source, or treat a nonempty stdout fragment as success. Syntax, formatting-directive, configuration, usage and I/O errors exit 2. Syntax, directive and configuration failures produce no formatted stdout. Status 1 belongs to --check, so use the default stdout mode for this recipe. --write writes disk files and does not format an unsaved buffer.

A generic adapter’s relevant operation is:

# buffer_bytes and destination_path come from the editor.
original = buffer_bytes
try:
    result = subprocess.run(
        [formatter_path, "--stdin-filepath", destination_path],
        input=original, capture_output=True, timeout=30, check=False,
    )
except (OSError, subprocess.TimeoutExpired) as error:
    show_process_error(error)             # keep original
else:
    if result.returncode == 0:
        replace_buffer(result.stdout)    # complete output, including empty output
    else:
        show_diagnostics(result.stderr)  # keep original

formatter_path is the built executable, not cargo run. The save command checks syntax and formats one buffer. It does not load includes, run semantic lint, invoke MiniZinc or discover files in the workspace. No editor extension, persistent process or formatting cache is required.

Measured behavior

The current bounded save checkpoint records the current release’s named cases. All expected outputs/statuses pass, but the default 966,669-byte dense case misses the 100 ms p95 gate at 111.014 ms. Overall interactive acceptance remains open pending a bounded dense-case follow-up. The historical results below remain separate.

Historical base-081 results

The save performance checkpoint records 50 fresh processes per case and configuration, complete stdout capture, error-buffer preservation and separate native child RSS checks. The original base-048 representative cases passed their budgets; the later public wide matrix prevented an overall acceptance claim.

The base-081 recheck uses the existing scalar-matrix renderer and syntax storage with no additional source change. Fifty fresh saves per configuration measure 89.184/89.121 ms p95 for the 885,733-byte public wide input. Separate native stdin saves measure 51.406250/51.375000 MiB for default/nested settings, passing the unchanged 100 ms / 64 MiB gate. Every retained representative case within 1 MiB passes its own latency/RSS budget. The dense 966,669-byte control measures 87.134/87.118 ms p95 and 47.796875/47.812500 MiB. Complete output and status match the retained baseline; focused matrix/table fidelity and scope-drop checks pass.

The checkpoint preserves the earlier failed campaigns, unsuccessful experiments, separate CPU/allocation evidence and first-use observations. Its first-use controls show no material launch delay, but do not explain the historical host startup cause or establish controlled cold-cache performance. Large inputs remain supported; interactive budgets cover inputs up to 1 MiB. Fresh full no-rules corpus validation remains deferred under the user’s explicit sweep restriction. The older corpus report has different source pins and cannot prove current full-corpus behavior. These save results do not resolve corpus or batch-memory limits.