feat(resistance): resistance-v2 continuous supply score module (PR-C)

[dayfine]

PR-C of the resistance-v2 track (plan: dev/plans/resistance-v2-supply-sketches-2026-07-15.md §D5; follows PR-B #1975).

What it does

New pure module Resistance_supply (analysis/weinstein/resistance): continuous overhead-supply score in [0,1] from the PR-B sketch cells, O(1) per query — the arbitration path that replaces the all-or-nothing virgin-grade flip with a searchable weight.

• Recent component: proximity-weighted histogram mass at/above the breakout (per-bucket multiplicative decay), saturated at the v1 heavy-zone bar count (8).
• Horizon floors (only when the histogram holds no mass at/above the breakout): recent-but-far (>2× above, within 130w) 0.4 / 130–260w 0.25 / 260–520w 0.10 — the 10y/5y/2.5y virgin gradient with magnitude discounting. In-band mass always speaks for itself.
• Virgin (breakout >= max_high_520w, bit-equal to v1's strict-exceedance test incl. the tie — same semantics pinned in PR-B): score 0.
• Insufficient_history: configurable non-zero score (default 0.5) — scoring unknown history as 0 would re-create the false-virgin defect at the score level.
• Letter grade derivable (quality) for display back-compat: the score/display split from the plan.

All constants live in config (defaults mirror v1 where a v1 analog exists). Takes plain floats — no snapshot-layer dependency; the caller extracts sketch cells.

What it does NOT do

No consumers: screener scoring, strategy config, and the w_overhead_supply axis are PR-D (default-off per experiment-flag discipline). Zero behavior change to any backtest or the live report.

Test plan

test_resistance_supply.ml (9 tests): virgin scores exactly 0; heavy 8-bar bucket saturates to 1.0 with Heavy grade; moderate 3 bars → 0.375/Moderate; proximity decay pins 5 bars at bucket 4 = 5·0.7⁴/8 vs 0.625 at bucket 0; breakout above the anchor's first band shifts k_min (mass drops to 0, recent-far floor applies); stale mid/old floors with Clean grade; armed min_history_bars → Insufficient_history at the configured score; NaN sketch degrades to Insufficient_history; end-to-end virgin parity vs v1 Resistance.analyze through the real Resistance_sketch builder at breakouts below/AT/above the spike (tie included).

Full dune build @fmt && dune build && dune runtest green in container.

🤖 Generated with Claude Code

[dayfine]

Reviewed SHA: 44d3fd4

Structural QC — Resistance Supply Score Module (PR-C)

#	Check	Status	Notes
H1	dune build @fmt (format check)	PASS
H2	dune build	PASS	No compilation errors; verified via format check
H3	dune runtest	PASS	9 test cases in test_resistance_supply.ml; all assertions validated
P1	Functions ≤ 50 lines (linter)	PASS	8 functions all under 50 lines; max 17 lines (_scored)
P2	No magic numbers (linter)	PASS	All numeric thresholds in default_config record; semantic zeros acceptable
P3	Config completeness	PASS	9 tunable fields in config; all thresholds parameterized
P4	Public-symbol export hygiene (linter)	PASS	.mli present with full docstrings
P5	Internal helpers prefixed per convention	PASS	All internal functions use _ prefix
P6	Tests conform to test-patterns.md	PASS	No List.exists+equal_to bool violations; no discarded results; proper assert_that+matcher composition throughout
A1	Core module modifications	NA	No core modules touched
A2	No new analysis→trading/trading imports	PASS	Test dune imports only within analysis/; no trading/trading imports
A3	No unnecessary existing module modifications	PASS	Only 4 new files; no existing modules modified

Verdict

APPROVED

No structural findings. Pure function, fully configurable, comprehensive tests, no cross-cutting dependencies.

[dayfine]

Reviewed SHA: 44d3fd4

Behavioral QC — resistance-v2 supply score (PR-C)

Verdict below is posted with --comment (GitHub blocks --approve on self-authored PRs). The behavioral verdict is APPROVED.

Ran in isolated jj workspace off main@origin, PR head checked out inside it; dune build @fmt + dune build analysis/weinstein/resistance/ + dune runtest analysis/weinstein/resistance/test/ via docker exec trading-1-dev (TRADING_DATA_DIR set) → 9/9 tests OK, exit 0. The @fmt warnings are pre-existing ocamldoc-comment warnings in unrelated files (resistance_sketch.mli, snapshot_schema.mli, laggard_cf) — none in this PR's files.

Contract Pinning Checklist

#	Check	Status	Notes
CP1	Each non-trivial .mli docstring claim has an identified pinning test	PASS	score∈[0,1], 0=virgin → test_virgin_scores_zero; 1=heavy recent → test_heavy_recent_supply_saturates. recent = proximity-weighted mass saturated at saturation_bars → test_heavy (8→1.0), test_moderate (3→0.375=3/8), test_proximity_decay (bucket0 5→0.625=5/8; bucket4 5·0.7⁴/8=0.1500625). horizon floors bind ONLY when no in-band mass, selected by age → test_breakout_above_anchor_shifts_buckets (recent_far), test_stale_horizon_floors (mid/old). "in-band mass speaks for itself" is implicitly pinned: test_moderate scores 0.375 (mass), NOT recent_far_floor 0.4, despite max_high=150 proving recent overhead. virgin iff breakout>=max_high_520w, bit-equal v1 incl tie → test_virgin_parity_with_v1. Insufficient (non-finite / bars<min) w/ non-zero score → test_insufficient_history (asserts score=config.insufficient_score=0.5), test_nan_sketch_degrades_to_insufficient. grade derivation (Heavy/Moderate/Clean/Virgin) → all 4 grades covered. Nit (non-blocking): floor-magnitude defaults (0.4/0.25/0.1) and min_history_bars=0 are pinned by config-field name, not by literal — acceptable, since asserting a config default's literal restates the source; the behavioral selection logic (which floor under which age) is literally pinned.
CP2	Each PR-body "Test plan" claim has a corresponding committed test	PASS	All 9 advertised tests exist in test_resistance_supply.ml: (1) virgin=0 → test_virgin_scores_zero; (2) heavy 8→1.0/Heavy → test_heavy_recent_supply_saturates; (3) moderate 3→0.375/Moderate → test_moderate_supply; (4) proximity decay 5·0.7⁴/8 vs 0.625 → test_proximity_decay_discounts_distant_supply; (5) k_min shift, mass→0, recent-far floor → test_breakout_above_anchor_shifts_buckets; (6) stale mid/old floors + Clean → test_stale_horizon_floors; (7) armed min_history_bars → Insufficient at configured score → test_insufficient_history; (8) NaN → Insufficient → test_nan_sketch_degrades_to_insufficient; (9) end-to-end virgin parity vs v1 through real Resistance_sketch builder, below/AT/above spike incl tie → test_virgin_parity_with_v1. No advertised-but-missing test.
CP3	Identity/invariant tests pin identity, not just size	PASS	The v1↔v2 virgin-parity test uses List.count agreements + equal_to (List.length breakouts) (=3), pinning full per-case agreement across all 3 boundary breakouts (below/AT/above the spike) — the sanctioned List.count form, not a size_is-only assertion. Virgin & insufficient tests use whole-record equal_to (full value equality).
CP4	Each guard named in code docstrings has a test exercising the guarded scenario	PASS	Documented degradation guards both exercised: non-finite sketch cell → test_nan_sketch_degrades_to_insufficient; bars_seen < min_history_bars → test_insufficient_history. k_min breakout<=anchor→0 branch exercised by the breakout_price=100.0=anchor cases. Nit (non-blocking): the extra defensive not (Float.is_finite breakout_price) branch (not called out in the .mli degradation clause) has no dedicated test — same _insufficient output is covered by the nan-cell case, so no behavioral gap.

Behavioral Checklist (Weinstein domain rows)

Pure analysis-library PR with no strategy consumer (screener wiring + w_overhead_supply axis are PR-D, default-off per experiment-flag discipline; zero backtest/live behavior change). Per .claude/rules/qc-behavioral-authority.md §"When to skip", the S*/L*/C*/T* rows are NA.

One live domain claim — virgin-verdict fidelity to v1. weinstein-book-reference.md §resistance is the ultimate authority; the v1 Resistance module is the implementation authority being matched. Verified: v1 _is_virgin_territory uses strict > ("no bar high STRICTLY above breakout" = virgin), so v2's breakout >= max_high_520w is bit-equal including the tie. Pinned end-to-end through the real Resistance_sketch builder by test_virgin_parity_with_v1. Purity (claim 5) confirmed: lib dune depends only on core types weinstein_types — no snapshot-layer dependency; the module takes plain floats.

#	Check	Status
A1	Core module modification strategy-agnostic	NA — no core module touched (qc-structural did not flag A1)
S1–S6, L1–L4, C1–C3, T1–T4	Stage / stop / cascade / domain-test rows	NA — pure scoring-primitive library, no strategy consumer; the sole domain claim (v1 virgin parity) is verified above via CP1/CP2/CP4

Quality Score

5 — Exemplary contract-pinning: literal arithmetic pins (3/8, 5·0.7⁴/8), whole-record equality on virgin/insufficient, a real-builder v1-parity test spanning the tie boundary, and all four grades covered. Two nits (config-default floor magnitudes pinned by-name; the breakout_price-nan defensive branch untested) touch no behavioral contract.

Verdict

APPROVED