docs: state unwritten contracts in doc comments

[cameroncuster]

Documents unwritten contracts — preconditions the code relies on but the //! doc comments didn't state. Comment-only change (0 code lines touched), plus two line-pin bumps in the cppcheck suppression list.

Each contract was verified against the exact line that breaks if it's violated:

File	Contract added	What happens if violated
graphs/dijkstra.hpp	requires weights >= 0	silently wrong distances
strings/binary_trie.hpp	requires 0 <= num < 2^61; walk requires a non-empty trie	wrong result (bits above 60 ignored; walk on empty trie follows -1)
math/mod_division.hpp	requires gcd(y, mod) == 1	assert fires
math/matrix_related/row_reduce.hpp	requires mod is prime	mod_div asserts / wrong answer (needs every nonzero pivot invertible)
math/matrix_related/matrix_mult.hpp	watch for overflow if T is an integer type	silent overflow (fine with modint, as in tests)
graphs/uncommon/enumerate_triangles.hpp	requires no self loops or multi edges	self-loop u==v makes tie comparison degenerate; multi-edges double-count
loops/submasks.hpp	excludes the empty submask	loop condition submask; stops before 0 — easy 4am surprise
math/derangements.hpp	requires n >= 1	dp[0] = 1 on empty vector is UB
strings/kmp.hpp	requires t non-empty	pi[j - 1] with j==sz(t)==0 reads pi[-1]
trees/hld.hpp	g is mutated (parent edges removed)	caller reuses g afterward and gets a broken graph
trees/centroid_decomp.hpp	g is mutated (edges removed)	same
trees/edge_cd.hpp	g is mutated (edges reordered/removed)	same

Findings that were considered and deliberately NOT documented (checked against the code):

• "trees must be rooted at 0 / connected" — universal convention across every tree structure in the repo; documenting it on 7 files individually adds noise, and vector<basic_string<int>> g(n) usage blocks already imply the shape
• "n >= 1" on DSU/SCC/rmq etc. — constructors are safe on n=0 (empty vectors, loops don't run); only ops on nonexistent indices break, which is true of every container
• rmq assert(l < r) — enforced by an assert already in the code, self-documenting
• suffix_array 0 <= s[i] < max_num — already documented in that header
• hungarian 1 <= n <= m — already documented (1<=i<=n<=m)
• several sub-agent claims were simply wrong (e.g. "partitions breaks on n=0" — vector dp(0) + non-executing loop is fine; "xor_convolution power-of-2 undocumented" — it's in the @code block)

Validation:

• diff outside the suppression list is 100% comment lines
• clang-format --dry-run --Werror passes on all changed files
• grep_clangformat_cppcheck.sh passes
• affected tests compile: dijkstra, derangement, matrix_mult, matrix_determinant (row_reduce+mod_division), kmp, binary_trie