Bump fork to latest

[gnurizen]

Moves the ebpf-profiler fork pin from 194f55bebd28 to 0b6ec2963d84. That's 281 fork commits, most of them parca-dev/opentelemetry-ebpf-profiler#341, which merged 50 rounds of upstream (195 commits, through open-telemetry/opentelemetry-ebpf-profiler#1906). The rest are fork PRs that landed on top of the merge.

Greatest hits

New unwinding and language support

• Python 3.14 tagged stack references (upstream *: Fix arm64 builds to choose correct BPF file #1807).
• Ruby JIT frames (upstream build(deps): update docker.io/library/busybox:1.35.0 docker digest to 224c1fd #1102).
• .NET dynamic method names, symbolized with cDAC (upstream build(deps): update module github.com/aquasecurity/libbpfgo to v0.4.9-libbpf-1.2.0 #1748).
• LuaJIT reads its offsets from the binary (upstream chore(deps): jsonnet-bundler Lock file maintenance #1648). This is our own LuaJIT work coming back from upstream. Fork profiler: Use a buffer pool for write profiles #344 also fixes statically linked luajit executables, like Debian/Ubuntu's /usr/bin/luajit, which previously never handed off to the LuaJIT unwinder at the outer C/Lua boundary.
• Node: the fork finds js_dispatch_table_ with upstream's asm interpreters instead of hand-written disassembly (fork profiler: Increase rlimit #345), so the leaptiering support can go upstream. The x64 path now has its first test.

Unwinder fixes

• x86-64 frame pointer recovery when the return address is in a register (upstream build(deps): update docker.io/library/busybox:1.36.1 docker digest to b58e890 #1751).
• A negative CFA offset counts as invalid instead of a stack root (upstream Endless Refresh loop with perf map on kubernetes #1881).
• Fixed the arm64 bpf_user_pt_regs layout (upstream chore(deps): update dependency kubernetes/minikube to v1.31.1 #1877).

Processes and containers

• PIDs from descendant PID namespaces, such as sidecar deployments, are translated (upstream chore(deps): update pre-commit hook lucas-c/pre-commit-hooks to v1.5.1 #1538, build(deps): update docker.io/library/busybox:1.36.1 docker digest to a992d7d #1801).
• Processes resync on MMAP events (upstream Could not find .text section spams error log #1857). Fork Initial try for .eh_frame/.debug_frame based stack unwinding using eBPF #346 uses this to find dlopened parcagpu libraries and deletes the rtld dlopen uprobe. That uprobe patched any binary exporting a dlopen symbol, and ClickHouse refuses to start when its own text segment fails a hash check. Once the agent had seen one ClickHouse process, every later one on the host failed to start.
• Uprobe links no longer leak on shutdown (upstream tests: anyStackContains should evaluate every stack #1806).

GPU

• GPU PC sampling works on 5.15 kernels again (fork Format our C code #342). The tail-call program was over the 5.15 verifier limit, and the loader quietly dropped it. A program that fails to load now fails the attach instead of being skipped.
• Processes that map the parcagpu shim at different paths all get attached (fork Dockerfile.dev: Add delve to dev container #348). Before, only processes using the same path as the first one seen got GPU profiling, which hit any container that mounted the shim somewhere else.

Hardening

• Bounds checks on interpreter metadata parsing for Python, V8, HotSpot and .NET, plus gopclntab, so malformed or hostile process memory can't push the parsers into huge reads or loops (upstream build(deps): update github.com/google/pprof digest to eeec1cb #1645, grpc: Fix codec issue causes kubernetes metadata problems #1729, pkg/process: Save whether mapping contains debuginfo to upload #1730, pkg/profiler/cpu: Prevent creating unnecessary closure #1731, *: Introduce lightweight LRU cache and use in hotpaths #1732, build(deps): update github.com/google/pprof digest to 200ffdc #1844).

Internals

• The new probes API (upstream unwinding/dwarf: Failed to unwind using native unwinder (eh_frame-based) on musl linked binaries #1658, fix(config): fix reloading config from Kubernetes ConfigMap #1735, build(deps): update docker.io/library/busybox:1.36.1 docker digest to 6e49438 #1753), with off-CPU profiling moved out of the tracer into the probes/offcpu plugin (upstream Refactor converter component #1734, debuginfo/manager: Exclude the time to wait for goroutines turn #1739). Groundwork for the memory profiling series (upstream chore(deps): update golangci-lint to v1.53.1 #1744, pkg/process: Filter mappings to only consider executable sections #1745) and custom probes.
• Go 1.26 (upstream Add flags to control perf event buffer event handling #1778). oomprof v0.1.6 -> v0.1.7 (fork bpf: Add some basic metadata #347).

Agent changes

Bump ebpf-profiler fork to the upstream merge

• Off-CPU profiling goes through the probes/offcpu extension. Upstream only exposes it as a collector extension factory, so enableOffCPU builds the extension and enables its probe on the tracer. The threshold conversion matches the old code.
• Behaviour change: --profiling-duration must now be greater than zero. Upstream's collector config validation requires positive reporter and monitor intervals (upstream chore(deps): update docker.io/otel/opentelemetry-collector-contrib docker tag to v0.80.0 #1784).
• Upstream build(deps): update module github.com/alecthomas/kong to v0.8.0 #1788 turns off all metric collection when no OTel meter is configured, and the agent doesn't configure one. The fork keeps metrics on whenever a reporter is set, so our eBPF and interpreter metrics keep reporting.

Shut down when the tracer fails
When the tracer hit an unrecoverable error and closed Done(), the agent kept running without collecting anything. The trace loop now watches Done(), runs the normal shutdown, and exits with status 1.