CMAverse Paired-mval Bootstrap

This skill captures a verified workflow for running CMAverse 4-way decomposition
where, within a single bootstrap process, each replicate computes both
mval=list(0) and mval=list(1) on the same bootstrap indices, and the
full cmest object (not a compact summary) is saved per mediator per group.

It is a domain companion to clauder-rstudio-workbench: the orchestration
(one RStudio session -> N child R workers -> poll -> merge gate) is provided by
the fan-out harness there. This skill supplies the CMAverse-specific contract:
what a worker must do, what counts as a valid result, and how to map a
monolithic g_formula script into independent per-mediator workers.

For ordinary descriptive statistics and publication-ready three-line Table 1
outputs, route to the sibling $comparegroups-guide skill instead.

First case study: g_formula_new4.7 (original script region 552:3561). The
skill is written to generalize to new4.8.*; always re-do the region
inventory for a new script version before trusting the old region map.

Why paired mval

Running M=0 and M=1 as two separate cmest(..., inference="bootstrap") calls
draws independent bootstrap samples, so the M=0 vs M=1 contrast is contaminated
by resampling noise. Paired mval fixes one set of indices per replicate:

replicate b:
  indices_b   <- bootstrap sample
  effect_m0_b <- statistic(data, indices_b, mval = list(0))
  effect_m1_b <- statistic(data, indices_b, mval = list(1))

The implementation wraps cmest()/boot() so the original mval=list(0)
region also produces M=1 inside the same bootstrap. Do not run the original
M=0 and M=1 regions separately and then claim "same bootstrap".

First reads

• Map a monolithic script into workers: region-mapping.md
• Worker responsibilities and the three-file contract: worker-contract.md
• What a valid cmest result must satisfy: validation-standard.md
• Choosing and proving concurrency: concurrency-and-resources.md
• Post-processing the saved objects into tables/plots: postprocessing.md
• Failure modes, prohibitions, native-PR direction: failure-modes-and-pr.md

Workflow

1. Inventory the target script's mediator regions (see region-mapping.md).

2. Generate a fan-out contract with the helper, then validate it:

   python scripts\make_worker_contract.py `
     --worker-file <paired_worker.R> `
     --output-root "<OUTPUT_ROOT>" --run-id <RUN_ID> --max-parallel 7 `
     --mediators msat_c12_2,msat_c12_3,msat_c1_2,msat_c21_2,msat_c21_3,msat_c2_2,msateco_c2_2 `
     --groups sy_female,sy_male --nboot 10 --seed 12345 `
     --out task.yaml
   clauder-workbench native-smoke start --task-key cmaverse_paired_mval_<RUN_ID> --session-name default --agent codex --require-raw-file
   # Run real Codex native MCP tools and save each visible tool result to a raw text file:
   #   list_sessions -> smoke_list_sessions.txt
   #   execute_r("cat('NATIVE_EXECUTE_OK pid=', Sys.getpid(), '\n')") -> smoke_execute_r.txt
   #   execute_r_async("cat('NATIVE_ASYNC_DONE pid=', Sys.getpid(), '\n')") -> smoke_execute_r_async.txt
   #   get_async_result(<JOB_ID>) -> smoke_get_async_result.txt
   clauder-workbench native-smoke record --task-key cmaverse_paired_mval_<RUN_ID> --step list_sessions --ok --session-name default --raw-file smoke_list_sessions.txt
   clauder-workbench native-smoke record --task-key cmaverse_paired_mval_<RUN_ID> --step execute_r --ok --marker NATIVE_EXECUTE_OK --pid <R_PID> --raw-file smoke_execute_r.txt
   clauder-workbench native-smoke record --task-key cmaverse_paired_mval_<RUN_ID> --step execute_r_async --ok --job-id <JOB_ID> --raw-file smoke_execute_r_async.txt
   clauder-workbench native-smoke record --task-key cmaverse_paired_mval_<RUN_ID> --step get_async_result --ok --job-id <JOB_ID> --marker NATIVE_ASYNC_DONE --raw-file smoke_get_async_result.txt
   clauder-workbench native-smoke complete --task-key cmaverse_paired_mval_<RUN_ID>
   clauder-workbench fanout-plan --contract task.yaml --parent-evidence <native_smoke_PASS.json>

3. Lint the worker, then smoke first (nboot=10), then run:

   clauder-workbench worker-lint --contract task.yaml
   clauder-workbench fanout-run --contract task.yaml --parent-evidence <native_smoke_PASS.json> --max-parallel 7 --first-artifact-timeout-min 15

   worker-lint (also run automatically inside fanout-plan/fanout-run)
   BLOCKs any worker whose code_file contains sink(. A sink()-wrapped
   worker keeps the detached Rterm alive after the model finishes, so the job
   never exits and the slot is never released. Log via cat() +
   flush.console() and write_state() only.

   fanout-run submits and polls the workers itself through the Python MCP
   stdio client (its evidence is stamped MCP_STDIO_OK). It does not drive
   the agent's native mcp__r_studio__ wrapper. If a task must be proven over the
   agent's native MCP transport, use the native path instead of fanout-run:
   fanout-plan (emit per-worker submit codes) → submit each via the native
   wrapper → async-guard register-job (record each real job_id) → fanout-poll
   → merge-gate. fanout-run --transport native-wrapper deliberately BLOCKs and
   points here, so it can never silently masquerade as a native submission.

   Concurrency. --max-parallel is the starting ceiling. By default it is
   also fixed for the whole run. Add --auto-scale to implement the
   best-practice §6 behavior: each poll cycle the harness samples system memory
   and, while memory stays under --memory-threshold (default 85%) and workers
   remain pending, raises concurrency by one up to the worker count (or
   --max-parallel-cap); when memory crosses the threshold it stops launching
   new workers and lets the in-flight ones drain (it never kills a running job).
   Every adjustment is recorded in the run's scale_log.

   clauder-workbench fanout-run --contract task.yaml --max-parallel 3 `
     --auto-scale --memory-threshold 85 --first-artifact-timeout-min 15

4. Validate every mediator x group result before claiming success:

   python scripts\cmaverse_validate.py --output-root "<RUN_DIR>" `
     --mediators msat_c12_2,msat_c12_3,msat_c1_2,msat_c21_2,msat_c21_3,msat_c2_2,msateco_c2_2 `
     --groups sy_female,sy_male
   clauder-workbench merge-gate --contract task.yaml --parent-evidence <native_smoke_PASS.json>

5. Scale to formal run only after a resource probe (concurrency-and-resources.md).
   Do not extrapolate nboot=10 optimal concurrency to nboot=1000.

Hard rules

• Save the full cmest object, never a compact result table or effect vector.
• Never overwrite the original R script; workers reuse the original M=0 region.
• Never delete a local RData before validation passes.
• Never put credentials (cloud-drive accounts, UK/UID, tokens) in any markdown,
  asset, schema, or contract. The optional durable-archive/upload step takes
  credentials only from the environment or an untracked config.
• Never resubmit a long worker just because one MCP poll dropped: check the
  durable files and the Rterm process first.
• Never start a CMAverse fan-out run without fresh native-smoke PASS evidence
  from the current agent native MCP tool layer. The generated contract sets
  requires_native_smoke: true by default; use --no-require-native-smoke only
  for diagnostic MCP-stdio experiments, not for formal completion claims. Formal
  runs must use --require-raw-file; v0.3.3 and later require all four record
  evidence ids to be chained into the final PASS and preserve raw output hashes.
  v0.3.4 and later also reject any downstream native_smoke parent evidence
  unless those four chained record ids are present and unique.
• Never wrap the worker in sink(). A sink()-wrapped worker keeps the
  detached Rterm alive after the computation finishes, so the job never exits and
  the fan-out slot is never released — recovery then needs a manual cancel after
  confirming the durable output. Log via cat() + flush.console() and the
  write_state() helper. This rule is enforced: clauder-workbench worker-lint
  (run automatically by fanout-plan/fanout-run) BLOCKs any worker containing
  sink(.