Promote audit fixes F-23, F-31, F-32 to master

[adequatelimited]

Promotes three audit remediations from `audit-fixes` to `master`.

Contents

• F-23 / [F-23] HIGH: read_tfile() returns VERROR (1) through size_t API — error indistinguishable from 1-record success #99 / fix(tfile): read_tfile() returns 0 on error instead of 1 (F-23, #99) #158 — `fix(tfile): return 0 on error from read_tfile() to disambiguate from 1-record read`
  • `read_tfile()` now returns `0` on `fopen()` failure instead of `VERROR` (==1), which was indistinguishable from a legitimate single-trailer read.
  • Wired up the pre-existing TODO in `send_found()` to check `count != NTFTX`.
  • Unit test added: `src/test/tfile-read-error.c`.
• F-31 / [F-31] MEDIUM: get_hash() leaks socket descriptor on early error return paths #117 / fix(network): close socket on every exit path in get_hash() (F-31, #117) #159 — `fix(network): close socket on every exit path in get_hash()`
  • Restructured with a single `CLEANUP:` label so `sock_close()` runs on every exit path (was leaking `np->sd` on `send_op` / `recv_tx` failure).
• F-32 / [F-32] HIGH: catchup() shared unsigned count across OMP threads — underflow and runaway loop risk #112 / fix(sync): close data race on catchup() count read (F-32, #112) #160 — `fix(sync): close data race on catchup() count read via volatile shadow`
  • Local `volatile word32 count = count_in` shadow closes the race between the unsynchronized reader and the `OMP_CRITICAL_`-guarded writers. Underflow claim analyzed and determined not reachable (see [F-32] HIGH: catchup() shared unsigned count across OMP threads — underflow and runaway loop risk #112 comment).

Test plan

• Per-PR test plans executed (build clean, unit tests pass, live sync validation where applicable).
• Full build clean under `-Wall -Werror -Wextra -Wpedantic -fopenmp`.
• No consensus-path behavioral changes; all three are correctness fixes on error/race paths.