feat: Landauer experiments 19-25 + README enrichment + preprint updates

Landauer Erasure Structure (exp_19 to exp_25):
- exp_19: SEC local structure (replaced theta correction)
- exp_20: PAC global structure (replaced golden decay)
- exp_21: Coherence threshold detection
- exp_22: Ratio invariants across cascade levels
- exp_23: Precision tightening validation
- exp_24: Cascade precision analysis
- exp_25: Full stack validation (3 runs)
- All results JSON included

Documentation:
- README.md enriched with updated overview
- UNIFIED_EVIDENCE.md updated with latest findings
- PREPRINT_UPDATE_PLAN.md expanded
- PAC Series papers (balance_constant_decomposition, structure_cost_of_erasure) updated
- Changelog entry for README enrichment

Author: Peter Groom

.changelog/20260214_085700_readme_enrichment.md

@@ -0,0 +1,32 @@
+# README Enrichment: Depth, Metaphor, Evidence
+
+**Date**: 2026-02-14 08:57
+**Type**: documentation
+
+## Summary
+Second pass on README rewrite. Elevated the hammer/glass metaphor from infodynamics.md to the opening hook, replaced the generic cross-domain section with a precise evidence table drawn from UNIFIED_EVIDENCE.md, and added the derivation chain, Ξ validation, and falsifiability conditions.
+
+## Changes
+
+### Changed
+- Opening tagline replaced with hammer/glass metaphor as accessible entry point
+- "Why This Framework Spans Every Domain" → "Two Axioms, One Derivation Chain" — leads with PAC/SEC definitions and derivation chain
+- Cross-domain findings table expanded from 8 rows to 14, with actual precision numbers (ppm, p-values, digit counts)
+- Added Ξ = γ + ln(φ) four-source validation table
+- Added explicit falsifiability conditions section
+- Recommended Starting Points reordered: infodynamics.md first, UNIFIED_EVIDENCE.md second
+- TOC updated to match new section headings
+
+### Added
+- Link to UNIFIED_EVIDENCE.md for complete derivation chain
+- Derivation chain code block (PAC → φ → ln(φ) → Ξ → Feigenbaum → Standard Model → Maxwell)
+
+## Details
+User feedback: the cleaned README was thin relative to the actual experimental depth (170+ experiments, 14 domains). The infodynamics.md hammer/glass metaphor was identified as the best accessible entry point. Content drawn primarily from UNIFIED_EVIDENCE.md (799 lines) and PREPRINT_UPDATE_PLAN.md (381 lines) which document the full derivation chain and statistical summaries.
+
+## Related
+- Previous session: initial README cleanup (465 → 296 lines)
+- This session: enrichment (296 → 317 lines, but vastly more substantive)
+- Source: foundational/docs/preprints/UNIFIED_EVIDENCE.md
+- Source: foundational/docs/preprints/PREPRINT_UPDATE_PLAN.md
+- Source: infodynamics.md (hammer/glass metaphor)

README.md

@@ -68,11 +68,11 @@ connecting it to hyperbolic geometry and the geometric mean between successive F
 
 At first glance, γ (from the harmonic series) and ln(φ) (from Fibonacci recursion) have no obvious relationship. Their sum Ξ = 1.0584... is not a listed constant in mathematical databases.
 
-The central claim of this paper is that these constants *must* add because they represent complementary costs of the same process: maintaining PAC conservation across scale transitions. Specifically:
-- **ln(φ)** is the cost of recursive structure (how much information each PAC level contributes)
-- **γ** is the cost of discrete-to-continuous regularisation (the overhead of mapping between countable and continuous domains)
+The central claim of this paper is that these constants add because they arise from complementary aspects of the same process: maintaining PAC conservation across scale transitions. Specifically:
+- **ln(φ)** is the cost of recursive structure (how much information each PAC level contributes — derived in Paper 1)
+- **γ** is *consistent with* the cost of discrete-to-continuous regularisation (it governs the Mertens product, which describes PAC-conserving prime sieving)
 
-Any system that exhibits both recursive conservation (PAC) and operates across discrete-to-continuous scale transitions should produce a balance point at their sum.
+The interpretation of γ as "discrete-to-continuous cost" is supported by its role in Mertens' theorem and its appearance at the Rule 110 order/disorder boundary, but it has not been proven from first principles. What *is* established is that both constants arise independently from PAC-conserving processes and converge to Ξ across four computational domains. The question of *why* they must add remains open (§10.1).
 
 ---
 
@@ -240,7 +240,7 @@ Experiment exp_03 (originally exp_29 of prime_growth_dynamics) analysed the deco
 | Component | Value | Share of Ξ | Role |
 |-----------|-------|-----------|------|
 | ln(φ) | 0.4812 | 45.5% | Structure (recursive PAC unit) |
-| γ | 0.5772 | 54.5% | Surplus (discrete↔continuous bridge) |
+| γ | 0.5772 | 54.5% | Surplus (consistent with discrete↔continuous bridge) |
 | **Ξ** | **1.0584** | **100%** | Combined balance constant |
 
 The surplus-to-structure ratio is approximately 1.200. The Rule 110 midpoint (the density at which the automaton transitions between ordered and disordered behaviour) is 0.574, which matches γ = 0.577 to within 0.56%.

foundational/docs/preprints/PACSeries/balance_constant_decomposition/paper.md

@@ -13,7 +13,7 @@
 
 Landauer's principle requires that erasing one bit dissipates at least $kT \ln 2$ of energy. The data processing inequality requires that information dispersed across multiple modes creates inter-mode correlations. This paper asks what these two established results, taken together, require to be true about the structure of an environment after information is erased into it.
 
-We model single-bit erasure into multi-mode thermal environments across five coupling topologies (single-mode, uniform, exponential, random, cascade) and measure the emergent correlational structure $\xi$. We find that: (1) every multi-mode topology produces new inter-mode correlations, confirming the theoretical expectation; (2) the structure is topological, invariant under temperature from 100K to 5000K; (3) cascade coupling, which mirrors physical heat dissipation, produces the most structure; and (4) the collapse efficiency ratio $A/(A+\xi)$ at default cascade parameters falls within ~2% of $\ln\phi = 0.4812...$, consistent with a broader cross-domain pattern in which structural boundaries cluster near $\phi$-family constants without parameter tuning.
+We model single-bit erasure into multi-mode thermal environments across five coupling topologies (single-mode, uniform, exponential, random, cascade) and measure the emergent correlational structure $\xi$. We find that: (1) every multi-mode topology produces new inter-mode correlations, confirming the theoretical expectation; (2) the structure is topological, invariant under temperature from 100K to 5000K; (3) cascade coupling, which mirrors physical heat dissipation, produces the most structure; and (4) the collapse efficiency ratio $A/(A+\xi)$ converges toward $\ln\phi = 0.4812...$, reaching 0.15% proximity at $N = 5 \times 10^6$ samples with Miller-Madow bias correction and thermally initialized (Boltzmann) environments.
 
 We further demonstrate that the thermal residual $\Theta$ re-injects as potential for subsequent erasure events, producing a self-sustaining cascade with 53× amplification over single events ($p = 2.75 \times 10^{-35}$). The cascade creates a natural temporal asymmetry: early moments are computationally dense, late moments are sparse, with a 69× difference ($p = 3.25 \times 10^{-5}$).
 
@@ -609,3 +609,28 @@ The ratio does not converge monotonically to $\ln\varphi$ as $\Theta \to 0$; ins
 **Verification**: The derivation predicts $\xi/A = (1 - \ln\varphi)/\ln\varphi = 1.078$. Exp_14 measured $\xi/A = 1.086$, a 0.76% discrepancy. The proximity is consistent with the broader pattern but does not establish convergence to arbitrary precision.
 
 **Status**: Derived and partially validated (exp_16). The mathematical structure is correct; the match to simulation is approximate (~2%). See Data/results/exp_16_ln_phi_derivation_20260209_111320.json.
+
+### 15.3 Precision Tightening and Full Stack Validation (Feb 2026)
+
+Experiments 23–25 addressed two questions: does the ~2% gap close with more samples, and does the full derivation chain hold end-to-end?
+
+**Precision result (exp_23)**: At $N = 5 \times 10^6$ Monte Carlo samples per seed with Miller-Madow bias correction:
+
+$$A/(A+\xi) = 0.4820 \quad (+0.15\% \text{ from } \ln\varphi)$$
+
+The gap narrows monotonically with sample size: ~2% at $N = 5 \times 10^5$, ~1.6% at $N = 2 \times 10^6$, ~0.15% at $N = 5 \times 10^6$. This is consistent with finite-sample bias rather than a fundamental offset.
+
+**Thermal initialization discovery (exp_25)**: The Boltzmann distribution of environment mode occupation is *required* for ln(φ) to emerge. Uniform 50:50 initialization yields $A/(A+\xi) \approx 0.33$. The partition ratio is a property of *physical* erasure at thermal equilibrium, not an arbitrary binary process. This narrows the claim: ln(φ) characterizes the erasure partition specifically in thermally equilibrated multi-mode environments.
+
+**Full stack validation (exp_25)**: All six layers of the derivation chain validated in a single experiment:
+
+| Layer | Test | Result | Status |
+|-------|------|--------|--------|
+| 1. Algebraic PAC | $\varphi^2 = \varphi + 1$ | $< 10^{-14}$ | PASS (exact) |
+| 2. SEC dynamics | Critical $\lambda^*$ → $1/\varphi$ fraction | 0.08% error | PASS |
+| 3. Landauer single-shot | $A/(A+\xi)$ vs $\ln\varphi$ (50 seeds × 2M) | 1.6%, ln(φ) in 2σ | PASS |
+| 4. Cascade | Θ re-injection, multi-generation | Amplification 1.2× | PASS |
+| 5. Gauge hierarchy | $\xi(SU(3)) > \xi(SU(2)) > \xi(U(1))$ | $p < 10^{-18}$ | PASS |
+| 6. Ξ composition | 4 analytic sources converge | CV = 0.05% | PASS |
+
+**Status**: Complete (exp_23, exp_25). The derivation chain from PAC axiom through gauge hierarchy holds end-to-end with no layer failing.