[Phase 1.2] Refactor build_netstats() to use g-computation predictions from joint model

[smjenness]

Context

Once a joint Poisson GLM is fit per layer (see #61), build_netstats() should use it via g-computation to produce internally-consistent target statistics.

Proposed approach

Step 1 — Build synthetic target population (N=10,000 default) from a single reference distribution. Attributes: age (continuous), age.grp, race, deg.casl, hiv2, etc. See #64 for post-stratification options.

Step 2 — Predict per-node expected degree using the joint model:

pred_deg <- predict(netparams\$main\$joint_model, newdata = synth_pop, type = 'response')

Step 3 — Aggregate target statistics by subsetting:

edges_target      <- sum(pred_deg) / 2
nf.age.grp[k]     <- sum(pred_deg[synth_pop\$age.grp == k])
nf.race[r]        <- sum(pred_deg[synth_pop\$race.cat.num == r])
nf.deg.casl[d]    <- sum(pred_deg[synth_pop\$deg.casl == d])
nf.diag.status[h] <- sum(pred_deg[synth_pop\$hiv2 == h])

All stats are internally consistent by construction: Σ_k nf.age.grp[k] = Σ_r nf.race[r] = 2 × edges_target.

Step 4 — Add method = c('marginal', 'joint_gcomp') argument; default 'marginal' for backward compatibility.

Tasks

• Add method argument to build_netstats(). Default 'marginal' (current behavior).
• Under method = 'joint_gcomp', use joint GLM predictions on synth population.
• Refactor each target stat computation (edges, nf.*, concurrent, degrange) to use predicted per-node degrees.
• Remove edges.avg flag when method = 'joint_gcomp' (inconsistency is resolved by construction).
• Warn clearly when method = 'marginal' is used that joint approach may be more appropriate.
• Regression tests: method = 'marginal' reproduces existing output byte-identically.

Acceptance criteria

• build_netstats(method = 'joint_gcomp') returns a netstats object with internally consistent targets:
  • Σ_k nodefactor_age.grp[k] == 2 × edges within floating-point tolerance.
  • Σ_r nodefactor_race[r] == 2 × edges similarly.
• build_netstats(method = 'marginal') produces the same output as the current function.
• Documentation flags that method = 'marginal' may be biased when target distribution differs from ARTnet sample.

Related

• Blocked by: [Phase 1.1] Refactor build_netparams() to fit joint Poisson GLM across attributes #61 (joint model fitting)
• Blocks: [Phase 1.4] Support post-stratification via user-supplied target population distribution #64 (post-stratification), [Phase 1.5] Validation suite: marginal vs joint_gcomp on ARTnet 2017-18 #65 (validation)