feat(rewards): allocate emissions by bounded per-repo emission_share

[RenzoMXD]

Closes #1215

Summary

Replaces the per-PR repo_weight_multiplier factor with per-repo emission share allocation at aggregation time. Every PR scores as if its repo contributes 1.0 to the multiplier chain; the final reward is divided proportionally inside each repository's emission_share × OSS_EMISSION_SHARE slice. Allocate-then-distribute, never multiply-then-cap.

Deliverables (mapped 1:1 to #1215)

D1 — RepositoryConfig.weight renamed to emission_share. Load-time validation enforces per-entry [0, 1] and registry sum ≤ 1.0. Strictly greater than 1.0 is rejected; less than 1.0 is valid and routes the remainder to recycle.

D2 — Each active repo receives exactly emission_share × pool. Single eligible PR claims the full slice; many PRs split it proportionally by earned_score. A high-throughput repo cannot exceed its slice; a low-throughput active repo cannot under-claim it.

D3 — master_repositories.json migrated in-place. Current values already satisfy the new invariant (sum = 0.71), so 0.29 routes to recycle as registry slack. entrius/oc-1 gets explicit issue_discovery_share: 0.0 per the issue body's canonical optimization-track example.

D4 — repo_weight_multiplier removed from per-PR scoring, ScoredPR, PullRequest, Issue, mirror adapters, CLI miner-score breakdown, and the storage layer. No code path multiplies by per-repo weight at PR granularity.

D5 / D6 — tests/validator/test_load_weights.py covers registry invariants (per-entry bounds, sum-≤-1.0 rejection, issue_discovery_share default 0.5). tests/validator/emissions/test_allocate.py adds 15 named tests covering:

• issue_discovery_share = 0 short-circuit (the oc-1 canonical case) - issue_discovery_share = 1 short-circuit (opposite direction)
• Within-repo spill in both directions (pr_side_spills_to_issue_side AND issue_side_spills_to_pr_side)
• Cross-repo linked issues attach to the issue's home repo, not the solving PR's home repo (D8 enforcement)
• Multi-repo cumulative slices for a single UID
• Registry slack at sum = 0.8 routes 18% to recycle without redistributing
• Full-registry round (sum = 1.0, all active) sums to 1.0 with zero recycle
• No-activity round recycles the entire OSS pool with treasury preserved

D7 — Atomic ship. Both legacy normalize.py modules deleted (no backward-compat path). New allocator has no optional fallback params. SQL persistence drops repo_weight_multiplier from BULK_UPSERT_PULL_REQUESTS and discovery_repo_weight_multiplier from BULK_UPSERT_ISSUES (columns no longer written; existing DB columns can be dropped via ad-hoc ALTER TABLE post-merge, mirroring the precedent from #1202).

D8 — issue_discovery_share: float = 0.5 added to RepositoryConfig. Defaults to 0.5 (even PR/issue split) per the issue body. entrius/oc-1 sets 0.0 explicitly. Allocation attaches to the issue's home repo via Issue.repository_full_name; cross-repo mechanics (solving-PR token cache, one-issue-per-PR canonical owner, anti-gaming gates) are unchanged.

D9 — constants.py collapses to:

• OSS_EMISSION_SHARE = 0.90 (combined scoring pool — replaces both old
  OSS_EMISSION_SHARE = 0.30 and ISSUE_DISCOVERY_EMISSION_SHARE = 0.10)
• ISSUES_TREASURY_EMISSION_SHARE = 0.10 (down from 0.15, still flat to UID 111)
• RECYCLE_EMISSION_SHARE removed — recycle becomes natural slack from unclaimed per-repo slices plus registry-level shortfall.

D10 — Recycle and split behavior tested explicitly (see test names in D5/D6 above). Per-round invariant OSS + treasury + recycle = 1.0 asserted in test_round_totals_sum_to_one_when_all_repos_active_and_registry_full.

D11 — Per-repo allocation lives in gittensor/validator/emissions/allocate.py. The emissions/ namespace sits alongside oss_contributions/ and issue_discovery/ so allocation is no longer modeled as an OSS-only concern. Within-repo sub-slices spill across the PR/issue split when exactly one side is empty; both empty recycles the repo slice.

Monetary policy shifts (flagged per D10)

This change carries two intentional policy shifts on top of the architectural one:

1. Recycle baseline removed. Today recycle is guaranteed 45% of emissions every round regardless of activity. Under this change recycle only fires from (a) repos with no eligible nonzero-scored activity on either side and (b) registry slack when Σ emission_share < 1.0. Round-totals invariant validated by test.

2. PR-vs-issue split inside the OSS pool. Today's effective split is 75:25 (30 OSS + 10 ID, combined). With OSS_EMISSION_SHARE = 0.90 combined and per-repo issue_discovery_share defaulting to 0.5, repos without an explicit override now split 50:50 within their slice.
   entrius/oc-1 opts out via issue_discovery_share = 0.0.

Other preserved semantics

Per-miner collateral from open PRs is preserved: total_collateral_score × OSS_EMISSION_SHARE is deducted from each UID's allocated reward in forward.apply_collateral_deductions, so open-PR spam still costs the miner emission share.

Test plan

• ruff check . — clean
• ruff format --check . — clean
• pyright — 0 errors
• pytest tests/ — 739 passing (existing 728 + 11 new in emissions/test_allocate.py)

[anderdc]

Issue #1215 deliverables missed: D1 (sum-overflow validation logs but does not raise — a registry summing > 1.0 silently loads), D7 (SQL columns dropped from INSERT/UPDATE in storage/queries.py and storage/repository.py with no migration script). Closing.

[RenzoMXD]

Hi, @anderdc
I can see your points. Can you reopen this PR so that I can update my code to completely finish the issue?