nx_dbg_step.nx

nx_dbg_step.nx -- THE STEPPING DEBUGGER (/compare/lang LN12, watch symbol dbg_step_over). THE RUNG (lang.plan): "Stepping debugger -- the doctor lane's DT1: break and step on our own binaries over the line table we emit (one contract, two pages)". LB5 shipped the line table (lm_stmt_stamp, statement granularity, readelf-validated). This is the debugger ON it. WHAT WAS REUSED, NOT REWRITTEN (nx_spendgate 2026-08-25 named all three): nx_addr2line_lib.nx -- the sovereign DWARF v5 line-program primitives (uleb/sleb/u32/u64le, nx_a2l_prog_off, and nx_a2l_func_at which already reads .debug_info subprograms). Every ULEB/SLEB byte in this file is decoded by THAT library's functions; there is still exactly one DWARF integer decoder. nx_bootcap.nx -- bootcap_machine: the SAME device set the boot rulers measure. rv64im_min_decoder.nx-- nx_rv64im_decode_kind + field extractors, for call/return classification. nx_lineconf_lib.nx -- the one line-anchored conf reader. nx_elf_read.nx -- ELF64 section walk. WHAT IS NEW HERE, AND WHY THE INCUMBENT COULD NOT DO IT: nx_a2l_run answers ONE question -- "which line is at address X" -- and to do it cheaply it keeps only the best row and STOPS AT THE FIRST end_sequence. A debugger needs the WHOLE table and it needs it BOTH WAYS: * file:line -> address (setting a breakpoint) <- the direction the incumbent cannot answer * address -> row index (where am I) * row i -> row i+1 (where does this statement END) <- the step primitive So this file runs the line program once into a two-pass-sized ROW TABLE, across ALL sequences. It also decodes THREE opcode classes nx_a2l_run refuses: SPECIAL opcodes (the compact bias form), DW_LNS_const_add_pc and DW_LNS_fixed_advance_pc, plus any other standard opcode via the header's own standard_opcode_lengths table. That is not scope creep: it is the difference between a table that works on our emitter's output and one that works on a line table from any DWARF v5 producer. Anything still undecodable sets decode_complete=0 and then RESOLUTION REFUSES WITH "UNPROVEN" RATHER THAN "NO SUCH LINE" -- a partial table that answers "absent" is a confident false negative, which is the one failure mode a breakpoint resolver must never have. STEPPING SEMANTICS (the actual rung). A statement step is NOT an instruction step: dbg_step_into -- single-step until the covering line-table ROW CHANGES. Entering a callee is a row change, so this stops on the callee's first statement. dbg_step_over -- the same, except while the call DEPTH is above the starting depth nothing is tested: the whole callee runs and the stop lands on the statement AFTER the call, in the caller. A `ret` past the starting frame is depth < 0, which is tested (that is gdb's behaviour at the end of a function, deliberately kept).