Harden joint g-computation target_pop path for forward-projection use

[smjenness]

Robustness fixes for the joint g-computation target_pop path, found while reviewing the ARTnet to ARTnetPredict forward-projection bridge. All changes are additive to the joint path. The legacy method = "existing" output is byte-identical to pre-refactor (backward-compat snapshot 3/3, R CMD check 0/0/0).

What changed

• build_netstats(method = "joint", duration.method = "joint_lm") no longer crashes when a target_pop has an empty age stratum. The synth-duration aggregator now falls back to the within-ARTnet byage durations per empty stratum instead of passing NA into dissolution_coefs. A restricted-age population (common in forward-projected MSM data, e.g. dropping the 55-64 group) previously aborted the build.
• target_pop data.frame degree/role/risk columns outside the fitted support are clamped with a warning, preserving the sum(nodefactor) = 2 * edges invariant. Out-of-support values (e.g. deg.casl = 4) previously broke it silently.
• Nodefactor tabulations span the full level set via factor(), so an unobserved level (e.g. an absent risk quintile in a target_pop) no longer recycles a per-level vector. Byte-identical when all levels are present.
• New optional seed argument to build_netstats for reproducibility of the stochastic sampling path. Defaults to NULL (no change in behavior).
• Documented the joint target_pop semantics: supplied deg.main/deg.casl are covariates (degrees are re-predicted by g-computation, not transported as counts), and diag.status affects predicted degree, not just nodefactor_diag.status.
• Fixed the stale scenario table in method_comparison.R / .md (it listed scenarios the harness no longer runs) and regenerated the report.
• Added inst/validation/forward_projection_bridge_decision.md: a memo on the Option A vs B estimand decision for the ARTnetPredict bridge.

Validation

• Backward-compat snapshot: ALL MATCH 3/3 for both default args and method = "existing" (tolerance 0).
• Full testthat suite passes, including 15 new regression tests in tests/testthat/test-target-pop-robustness.R (crash, clamping, no-recycle, seed reproducibility).
• R CMD check: 0 errors / 0 warnings / 0 notes.

Notes

• Version bumped to 2.9.1 so downstream consumers (ARTnetPredict) can pin ARTnet >= 2.9.1. Drop that commit if you prefer a different versioning scheme.
• Docs were regenerated with roxygen2 8.0.0 (the repo was last built with 7.3.3), which accounts for the DESCRIPTION roxygen-key and package-Rd churn.
• Does not address the length-bias / truncation sampling concerns in Length-biased sampling and 5-partnership truncation bias in formation target stats #72; that remains open and orthogonal.