Known issues

Defects found in review and benchmarking, with the PR that fixed each. All fifteen are fixed and have regression tests; the numbers are referenced from code comments and tests (KNOWN_ISSUES #n). Current overhead figures are in bench/RESULTS.md; planned work is in ROADMAP.md.

#	Severity	Area	Problem	Fix
1	High	correctness	Inter-process counter double-counted when publisher and subscriber shared a process	#4
2	Medium	correctness	"Node liveness" was "ever initialized": nodes were never removed	#7
3	Medium	efficiency	Unresolvable callbacks (timers, services) took the write lock on every call	#2
4	Medium	robustness	Default output file was shared by all processes, so writes interleaved	#6
5	Low	robustness	std::stod on a bad env var threw out of a tracepoint and terminated the host	#5
6	Low	docs	Dangling issue reference; the rclcpp#2911 premise needed a distro caveat	#1
7	Low	behaviour	Idle topics printed TOPIC /x 0.000000 every window	#3
8	High	correctness	First window reported up to 2x the rate (timer first-fire off by one, nominal instead of measured window length)	#9
9	Medium	robustness	Exit-time use after free: tracepoints could run during static destruction	#10
10	Medium	robustness	fork() without exec: the child kept counting but never flushed	#11
11	Low	behaviour	Output file was append-only and unbounded while the README said "rolling"	#13
12	Medium	behaviour	A stalled subscription emitted no RECV line, so a stall looked like absence	#12
13	Low	efficiency	Thread-local cache was single-entry; the "0.2 ns/op" figure was the all-hits best case	#14
14	Low	build	The probe exported every symbol, not just ros_trace_*; lint deps were declared but unused	#15
15	Medium	efficiency	TLS cache hash aliased on allocator strides (about 90 % misses under realistic farms, +2-4 % workload CPU)	#19

Sub-items referenced from code: 8a is the timer first-fire off-by-one, 8b is the nominal window denominator (both fixed in #9).