unified and cleaned up CI scripts - #6

…lang/Linux requires -fuse-ld=lld)

…g (it seems to expect unsigned, whereas we are passing size_t)

… system depends on it

…idy-18 does not understand source-filter option

…are meant to be private

…al namespace

…rameter identifiers like l and I

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Inserts 20-byte keys (> default HEAPBOUND of 17) so the key suffix is
heap-allocated via new_uchar and its raw pointer embedded in the node
bucket.  Running under ASAN confirms the destructor frees every such
allocation without leaking.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

New Node At Depth One: exercises ptrie_internal.hpp:1252 (node = new
node_t{}) with byte==1 and base==fwd.  Six 128-byte keys fill a leaf
whose split_fwd sets fwd_n->_children[0..127] to self-loops; the 7th
insert (5 bytes, nitemsize=4 < HEAPBOUND) finds such a self-loop at
depth 1, triggering the base==fwd path and allocating the new node.

Split Node Both Partitions: exercises ptrie_internal.hpp:1154 (h_node =
new node_t{}) when both low and high partitions of a split_node are
non-empty.  Five 64-byte keys plus one 4-byte key fill a leaf; after
split_fwd the recursive split_node at p_byte=1 finds f[1]=0x40 for the
long keys (HIGH) and f[1]=0x04 for the short key (LOW), creating h_node.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

write_data admits a symbolic path where size=0 after (stored_uint16 >> 8),
making both pos<size guards false and leaving the returned key uninitialized.
Remove the suppressing key_t{} zero-init in operator* and add an iterator
value-correctness test that drives 200 ints through write_data (forcing fwd
splits so ps≥1 nodes are exercised) and verifies each dereferenced value via
set.exists() under UBSAN.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Reverts the uninitialized key_t key; back to auto key = key_t{};
so the clang-analyzer-core.uninitialized.UndefReturn false positive is
suppressed by construction: even if write_data somehow exits without
writing all bytes the return value is still defined.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

clang-tidy flags node->_count + 1 as a garbage-value operation because its
symbolic model does not track the in-class initializer `uint16_t _count = 0`
on node_t, treating new node_t{} as leaving _count uninitialised.  The new
test drives the base==fwd insertion path (first insert into an empty set),
verifies correct insert/exists/duplicate semantics, and lets ASAN/UBSAN
confirm that _count is properly zero-initialised on every fresh-node alloc.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

_count garbage value (line 1293):
  The analyzer does not track the in-class initializer
  `uint16_t _count = 0` on node_t, so it models new node_t{} as leaving
  _count uninitialized.  Add an explicit `node->_count = 0` assignment
  immediately after the allocation so the initialisation is visible to
  the symbolic model.

dest potential leak (line 1401):
  The heap key suffix allocated by new_uchar is embedded into nbucket via
  mem_store (a byte-level copy).  The analyzer cannot follow the pointer
  through this type-punning operation and reports it as potentially leaked
  when nbucket is moved into node->_data.  Suppress with NOLINTNEXTLINE at
  the diagnostic site; the pointer is freed by delete_uchar during cleanup.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

as clang-tidy seems to require files which are generated during the actual build

…n process

…4.04)

… reports on other libraries

Brought the build/test/lint/CI docs in line with the ci-scripts branch:
the merged tidy workflow, GCC-14/Clang-20 toolchains, release-deb presets,
heaptrack test variants, and example/ consumption checks.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Instruments the ptrie test targets and adds a `coverage` target that
produces an HTML + textual report: gcov/lcov for GCC, llvm-cov for Clang.
lcov >= 2.0 is handled via --ignore-errors unused so the never-matched
benchmark exclude pattern does not abort the report.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Added a "Code Coverage" section to README.md (build steps, tooling per
compiler, required packages) and moved the llvm-cov "mismatched data"
explanation there from cmake/coverage.cmake. Also pass
geninfo_unexecuted_blocks=1 to lcov capture so inlined/templated code
no longer triggers "unexecuted block ... with non-zero hit count".

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

"Erase Missing Keys" exercises erase() on absent keys (empty trie, never
inserted, double erase, prefix-sharing longer key), closing the only real
coverage gap among the deletion functions. Two further cases ("Erase Keeps
Dense Node", "Erase Deep Subtree Collapse") add partial- and deep-deletion
regression coverage.

The remaining uncovered lines in merge_regular/merge_empty are unreachable
defensive guards: merge_down already returns on the same count>SPLITBOUND/3
check before calling merge_regular, the differing-sibling-type branch is
preceded by a block that always returns, and merge_empty's upward-collapse
loop only runs at the final root-level collapse.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Adding tests that copy tries with heap-stored (long) and multi-entry inline
keys uncovered three bugs in node_t::clone, all in branches no prior test
exercised:
  - inline copy used an end iterator of data()+(lencsize-bdepth) without the
    per-entry offset, dropping every entry after the first;
  - the deep heap branch copied `encsize` bytes into an `encsize-bdepth`
    buffer (heap overflow);
  - the shallow heap branch accessed the stored pointer via as_array at an
    unaligned offset (UBSan misalignment).
The heap branches now use mem_load/mem_store like the rest of the code.

New cases (Copy Heap Keys Shallow/Deep, Copy Inline Keys) cover these paths
and assert deep-copy independence; the full suite passes under ASAN+UBSAN+SSP.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Forces a shared first byte to split fully and advance to the next byte while
suffixes are heap-stored, so split_fwd's move_data must migrate heap pointers.
Acts as a deterministic regression guard for that path, which existing random
split tests only exercise incidentally (and which lcov under-reports because
move_data is a generic lambda). Passes under ASAN+UBSAN+SSP.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…idy errs on g++-14 options

…s on modmap files that clang-tidy cannot find