Separate CSV output per run (run id + registry)

[Hackatro]

Problem

Melodie appends to existing CSVs and writes the header only when the file is new (MelodieInfra/table/table_objects.py, Table.to_file):

is_new_file = True if not os.path.exists(file_name) else False
writer = TableWriter(file_name, append=not is_new_file).write()
if is_new_file:
    writer.send(headers)

The output CSVs already carry id_scenario, id_run, period. But id_run is Melodie's run_num within a single invocation — always 0 here. Running the simulation twice writes two complete sets of rows into the same files under identical keys, with nothing distinguishing them.

export_bundle.py filters on id_scenario only:

edf = edf[edf["id_scenario"] == scenario].sort_values("period")

Result: a bundle with 730 rows for 365 periods, two simulations interleaved and sorted together. No error is raised. The only thing keeping the current output correct is deleting data/output/ by hand between runs.

Related: because the header is written only for a new file, a run whose columns differ appends rows under the old header, silently misaligned.

Decision — one directory per run

Config.output_tables_path() returns output_folder, which main.py sets, so this needs no Melodie internals.

data/output/
  runs.json
  a3f9c1e2-.../
    Result_Simulator_Environment.csv
    Result_Simulator_BrazilFarms.csv
  7b21d0f4-.../

• UUID generated once at startup, used as the directory name
• The same UUID becomes the Postgres run_id in the follow-up ticket, so both paths share one identity
• Fresh files per run make schema drift and PDL drift impossible
• Deleting a bad run is deleting a folder

runs.json provides the ordering the UUID does not: run_id, started_at, PDL filename and content hash, scenario ids, period_num, git sha, optional label.

Scope

• main.py — generate the run UUID, point the Melodie Config output folder at it
• manifest writer — append the run entry to runs.json
• export_bundle.py — --run <id> defaulting to the newest entry in runs.json; it resolves to the existing --input directory, --scenario unchanged
• README — document the layout

Out of scope: frontend run selection, Postgres.

Acceptance

• Two consecutive runs produce two directories and two runs.json entries
• Neither run's CSVs contain rows from the other
• export_bundle.py with no --run exports the newest run
• export_bundle.py --run <older id> exports that run
• No manual deletion of data/output/ is needed for correct output