Land the Dynamics graph, accounting, identity and history stack (#420, #436–#448)

[MaxGhenis]

Lands the thirteen green draft PRs from the September 9–22 Dynamics lane onto current master in one merge-commit integration, so later work builds on master instead of on stacked drafts. Each PR's head is merged unchanged (merge commits, no squash), so every original commit stays in history.

PR	Adds
#420	Optional mortality and ageing population graph (populace_dynamics.graph)
#436	Annual mortality graph trajectories with isolated evaluation
#440	Stock-flow accounting on annual mortality trajectories
#437	Claiming state preserved through disability conversion (engine/claiming.py)
#438	Experimental entrant schedules with explicit support boundaries
#439	Opt-in annual population stock-flow accounting (engine/accounting.py)
#441	Consecutive annual accounting-history validation
#443	Source person IDs preserved with append-only private keys
#444	Annual labor and covered-wage histories
#445	Mortality outcomes retained with source identities
#446	Annual earnings histories retained through mortality
#447	Original-cohort histories retained from assembled projections
#448	Compact cohort-history storage without repeated identity maps

Conflict resolution

The PRs conflicted only on four shared files. Each PR appended its own new modules to the birth-evidence reducer's POST_REVIEW_SOURCE_EXCLUSIONS and bumped the tier manifest independently.

• scripts/first_estimates_birth_evidence.py: union of all 20 new source paths; duplicate comments collapsed.
• tests/estimates/test_birth_evidence_artifact.py: the pinned tuple matches the script, and the reachability test also guards the seven identity/history modules alongside the PSID, graph, accounting, claiming-successor and entrant exclusions. No registered evidence pin changes.
• tests/tier_counts.json / tests/README-tiers.md: recollected on the combined tree (unit 2,207; artifact 2,672; integration_psid 848; reproduction_legacy 520; oracle_policyengine 159; total 6,406).

All other files merged without conflict. No gate, contract, amendment or runs/ artifact is touched.

Verification

• Each source PR passed its full CI on its own head (4 test shards, aggregate, lint, build; graph integration on 3.13/3.14 where applicable).
• This branch: full remote CI below; local unit tier run in parallel.
• Downstream use: the r3 all-real-core rehearsal (all eight projection callbacks, 2014–22) ran on Retain original-cohort histories from assembled projections #447's source. PR Store cohort histories without repeating identity maps #448's compact codec then round-tripped its 12 retained histories exactly (50,431 → 39,740 bytes; 17/17 equality checks).

Cross-family review and fixes

Five Claude reviewers read the stack's source cluster by cluster; each finding then went to two independent refuters. Of 13 reported findings, 11 were confirmed (2 medium, 9 low) and 2 rejected. All 11 are fixed with regression tests, each cluster independently verified:

• Graph (Add optional mortality and ageing population graph #420/Add annual mortality graph trajectories with isolated evaluation #436/Add stock-flow accounting to annual mortality trajectories #440): the one-year graph exported next-period rows carrying the prior year's survives=True and death_probability (no draw had happened); those rows now carry null outcomes and the evaluation refuses admitted rows that claim one. Out-of-int64 IDs and years are refused instead of wrapping (a wrapped year slipped past the fit cutoff). String and boolean weights, exposures and deaths are refused instead of coerced.
• Accounting (Add opt-in annual population stock-flow accounting #439/Validate consecutive annual population accounting histories #441): PopulationReconciliationError and HistoryAccountingError are now picklable and copyable with every field intact.
• Entrants (Recover experimental entrant schedules with explicit support boundaries #438): entrant_schedule.validate_projection_allocator refuses projection metadata whose allocator can hand a later birth an existing unreserved ID (a schedule built from a second allocator at the same start made one ID name both a dead entrant and a newborn). The check lives outside engine/loop.py because the loop is inside the birth-evidence reducer's reviewed-implementation identity. A test pins that the unvalidated loop still reuses the ID, so enforcing it in the loop needs a deliberate re-pin. Numpy activation years are normalized so audit records serialize.
• Identity (Preserve source person IDs with append-only private keys #443/Preserve annual labor and covered-wage histories #444): PersonIdentityMap (and the histories that embed it) are picklable. The recorder's native-versus-private key hazard is documented, not guarded: nonnegative native IDs below the map size are indistinguishable from private keys, so the recorder cannot detect the mistake. The wired Retain original-cohort histories from assembled projections #447 path translates keys correctly.
• Histories (Retain original-cohort histories from assembled projections #447): float audit years are refused (one observation could otherwise validate under two digests), and numpy draw indexes and scheduled-year keys are normalized.

The review also caught a duplicated test that my union merge introduced (removed). Local unit tier after the fixes: 2,167 passed, 131 skipped (graph tests needing the graph extra, covered by CI's graph-integration job), 0 failed. Tier manifest: unit 2,298, total 6,497.

Closes #420, #436, #437, #438, #439, #440, #441, #443, #444, #445, #446, #447, #448 on merge.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Sep 22, 2026 9:49pm UTC