Synth-aggregated duration g-computation in build_netstats (#73)

[smjenness]

Closes the partial task on #63 raised by PR #71's review: under method = "joint" + duration.method = "joint_lm" in build_netparams, the within-ARTnet stratum aggregation that build_netparams emits is now overridden in build_netstats with synth-population aggregation.

What changed

• New private helper .aggregate_synth_byage_durations() in R/NetStats.R.
  • Predicts joint_lm log-duration at same.race = 0 and 1 per synth ego.
  • Marginalizes over partner-race uncertainty using P(same.race | ego) from joint_nm_race_model.
  • Exponentiates, takes median per stratum, applies the existing geometric transformation mean.dur.adj = 1 / (1 - 2^(-1/median)).
• build_netstats joint synth-prediction block now aliases synth$index.age.grp <- synth$age.grp (joint_lm uses index.age.grp RHS naming; joint_nm_*_model uses age.grp) and calls the helper for main and casl. Skips when joint_duration_model is NULL.
• diss.byage dissolution_coefs calls read from a local override vector when present, else fall back to netparams[[layer]]$durs.<layer>.byage$mean.dur.adj.
• diss.homog still uses the within-ARTnet aggregation (not consumed by EpiModelHIV-Template's tergm offset; can land in a follow-up).

sex.cess.mod handling: helper preserves the deterministic post-cessation "dead row" (mean.dur.adj = 1) at the end of the override vector so dissolution_coefs sees the same shape it did before.

Empirical effect

Stratum-level mean.dur.adj (Atlanta + race = TRUE, N = 10k):

Layer/Stratum	ARTnet-cond. (PR #71)	synth-Atlanta (this PR)	synth-NHBS shift
main / nonmatch	235.5	243.0	234.2
main / matched.5	445.6	491.5	491.5
casl / nonmatch	92.3	86.9	86.9
casl / matched.1	48.9	68.9	47.9
casl / matched.5	129.7	186.0	126.9

Main durations move modestly (joint_lm has weak race effects on duration); casl moves substantially under shifted population (matched.1: 68.9 Atlanta vs 47.9 NHBS) — exactly the marginal-vs-joint correction the refactor was designed to apply. This is the dyad-level analog of the formation-stat divergence we documented in PR #68 / #69 reviews.

Validation

• Backward-compat snapshot harness: 3/3 match on default and explicit method = "existing".
• New unit tests: tests/testthat/test-duration-gcomp-synth.R, 7 blocks, 17 assertions. Covers: override fires under joint + joint_lm; falls back under empirical durations; falls back under method = "existing"; diverges under shifted race.prop; preserves sex.cess.mod dead row; produces well-formed disscoef objects; handles race = FALSE without joint_nm_race_model.
• Full testthat suite: 525 / 525 pass.
• R CMD check: 0 errors / 0 warnings / 0 notes.
• End-to-end EpiModelHIV-Template: all 6 ERGMs (3 layers × default + joint methods) converge cleanly under Stochastic-Approximation. Main coef.form shifts -17.570 → -17.602 between PR Duration methods: empirical + joint_lm (#63 phase 3) #71's path and this PR's, consistent with matched.5 mean.dur.adj moving 445.6 → 491.5 under synth aggregation.

Approach note

This implements Option A from #63: ego attributes only on the prediction RHS, with partner-race marginalized via the existing joint_nm_race_model. No explicit synthetic partnership-pair construction (which would be Option B). The simpler choice is consistent with how nodematch/absdiff are handled in PR #69 and gives correct results for the cross-sectional-age-of-extant-ties target Steve Goodreau articulated in the PR #71 review.

Test plan

• Backward-compat snapshot 3/3 (default + explicit method = "existing")
• Override fires under joint + joint_lm; falls back under empirical / method = "existing"
• Casl durations diverge >1% under shifted race.prop on at least one stratum
• sex.cess.mod dead row preserved
• race = FALSE path works
• Unit tests 525/525
• R CMD check 0/0/0
• End-to-end EpiModelHIV-Template estimation under both default and joint paths converges

Depends on #71 (merged). Part of #63 — closes the partial task. Unblocks #65 (Phase 1.5 validation suite).