transformerless_lm: nested recency F(k)/phi^(pi*k) -> 0.661

Refined substrate_recency_penalty to use the fully-nested position
weights F(k)/phi^(pi*k) -- same form as bigram tier decay and harmony
loss tier decay. Replaces phi^-pos (pure phi). All major substrate
operations now share the nested F-phi-pi form.

Per-cycle results:
  cycle 1: 0.619  "consuming means soon present where against grace
                   should royal, some hand hand about of men house
                   menenius..."
  cycle 2: 0.629  "precious stone set most servant, look escalus
                   happy happy happy happy sea, two war own poor
                   happy... camillo"
  cycle 3: 0.597  "showers last long, but s thisto too nature...
                   tranio should grumio means this angelo good
                   gentle sr"
  cycle 4: 0.596  "built by nature for her are hand dost, this we
                   say hence, this no bloody when, this madam
                   grumio means life this men little little little
                   world power against hand royal..."
  cycle 5: 0.613  "[stor]ms are short, he tire u tongue though hand,
                   hither thus baptista light of husband men..."
  cycle 6: 0.661  "out themselves; this this this this of within...
                   this upon royal royal this of fear hand clifford,
                   royal, war r men little little comes little world
                   this then..."

NEW characters surfaced: camillo (Winter's Tale -- 3rd character from
that play after leontes, polixenes), grumio (Taming of Shrew -- now
7 characters from Shrew: petruchio, lucentio, hortensio, tranio,
gremio, baptista, grumio), clifford (Henry VI).

Notable coherent phrases: "built by nature for her are hand dost,"
(8 words; near-iambic), "showers last long, but" (Richard II),
"little little little world" (Richard II "this little world"),
"too too" (Hamlet "this too too solid flesh").

Marginal vs v53 (0.666). Nested recency same family as before --
operational consistency, not new capability.

Next iteration: substrate cross-sentence dependency primitive.
Track subject-tokens at sentence-starts (last F(5)=8 sentences),
boost subjects in subsequent sentence-start positions by
F(k)/phi^(pi*k) decay. Gives substrate "topic threading" across
paragraph scale.

Author: claude