feat(scoring): drop density, swap base score to exponential saturation

[seroperson]

Summary

Replaces the density-based base score with an exponential saturation curve on source token score, per the discussion and agreed shape in #1339:

base_score = MERGED_PR_BASE_SCORE * (1 - exp(-src_tok / SRC_TOK_SATURATION_SCALE))
           + min(total_score / CONTRIBUTION_SCORE_FOR_FULL_BONUS, 1) * MAX_CONTRIBUTION_BONUS

The new curve is monotonic in src_tok, concave from the origin, and ruled by one knob (SRC_TOK_SATURATION_SCALE = 58, per-repo overridable).

Per-repo override range is [10, 500], enforced by the loader and pinned by the live-config test in tests/validator/test_load_weights.py.

Requirements

#	Requirement	Status
1	Ship formula base_score = 25 * (1 - exp(-src_tok / 58)) + min(total_score / 1500, 1) * 5	✅
2	Drop contribution bonus cap to 5 instead of removing the term (keep tests/non-code non-zero, 5:1 source-vs-rest ratio)	✅
3	Land formula + bonus change in the same PR	✅
4	Delete MAX_CODE_DENSITY_MULTIPLIER	✅
5	Delete MIN_TOKEN_SCORE_FOR_BASE_SCORE from scoring.py (the cliff)	✅
6	Delete MIN_TOKEN_SCORE_FOR_BASE_SCORE from credibility.py (valid_merged_count filter)	✅
7	Eligibility becomes a pure count of merged PRs, per-repo overridable	✅
8	Add SRC_TOK_SATURATION_SCALE = 58	✅
9	Make SRC_TOK_SATURATION_SCALE per-repo configurable via master_repositories.json scoring block	✅
10	MAX_CONTRIBUTION_BONUS: 25 → 5	✅
11	Keep MERGED_PR_BASE_SCORE = 25	✅
12	Keep CONTRIBUTION_SCORE_FOR_FULL_BONUS = 1500	✅
13	Keep TEST_FILE_CONTRIBUTION_WEIGHT = 0.05	✅
14	Keep the full multiplier chain untouched	✅
15	Enforce [10, 500] range for SRC_TOK_SATURATION_SCALE via the load weights test suite	✅

Behavior changes

New score bounds:

• initial_base_score: 28.75 -> 25
• Total base_score: 53.75 -> 30

Also:

• Contribution bonus cap: MAX_CONTRIBUTION_BONUS 25 -> 5.
• Eligibility (check_eligibility) is now a pure count of merged PRs; per-repo min_token_score_for_base_score field removed.
• _build_solving_pr_cache no longer pre-filters by token score - every merged PR is reusable for issue-discovery lookups.
• PullRequest.code_density / ScoredMirrorPR.code_density default flips 0.0 -> 1.0. The field is no longer populated by the scorer (density is gone), but the column survives in the DB and is read by downstream consumers as a neutral multiplier - 1.0 keeps it neutral, 0.0 would zero anything multiplying by it.

Constants

Constant	Before	After
MAX_CODE_DENSITY_MULTIPLIER	1.15	removed
MIN_TOKEN_SCORE_FOR_BASE_SCORE	5	removed
MAX_CONTRIBUTION_BONUS	25	5
SRC_TOK_SATURATION_SCALE	n/a	58.0 (per-repo, [10, 500])
MERGED_PR_BASE_SCORE	25	25
CONTRIBUTION_SCORE_FOR_FULL_BONUS	1500	1500
TEST_FILE_CONTRIBUTION_WEIGHT	0.05	0.05

The full multiplier chain (time decay, review quality, label, issue, spam, credibility) is untouched.

Per-repo src_tok_saturation_scale

Optional knob on each master_repositories.json scoring block:

"scoring": { "src_tok_saturation_scale": 80.0 }

Loader rejects values outside [10, 500]. No entry in master_repositories.json sets the knob today; default is SRC_TOK_SATURATION_SCALE = 58.0.

Open concerns / follow-ups

• code_density column lingers in the DB. Density is gone from scoring, but the column survives on pull_requests (written by the upsert in gittensor/validator/storage/queries.py) and the field survives on PullRequest / ScoredMirrorPR. The default flip to 1.0 keeps downstream consumers safe, but I'm unsure how DB schema removal is handled in this repo. Open question: schema-level removal as a follow-up PR, or leave the column as a permanent neutral 1.0 write?
• Eligibility no longer requires source-code change. The saturation curve suppresses score for trivial PRs but not count toward MIN_VALID_MERGED_PRS = 3. A miner can clear the gate with three test-only / non-code merged PRs. Per direction (comment), eligibility is intentionally "a pure count of 3 merged PRs", but still worth tracking: should there be maybe an opt-in require_source_for_eligibility per-repo knob so test-only/non-code-only PRs count toward credibility but not the minimum-merged gate?

Related Issues

Closes #1339.

Type of Change

• New feature
• Bug fix
• Refactor
• Documentation
• Other (describe below)

Testing

• Tests added/updated
• Manually tested using gitt miner score

Checklist

• Code follows project style guidelines
• Self-review completed
• Changes are documented (gittensor-docs-ui update to follow on test)

Test plan

uv run pre-commit run --all-files
uv run pre-commit run --all-files --hook-stage pre-push

New coverage:

• TestPerRepoSaturationScaleOverride - default scale + per-repo override reshapes the curve in both directions.
• test_loader_rejects_out_of_range_saturation_scale + test_loader_accepts_saturation_scale_at_bounds - parametrized [10, 500] bounds.
• test_live_mirror_scoring_fields_have_valid_shape - live master_repositories.json pinned to the same bound for every repo.