size_t i = strlen(s) - 1: the (i > 0) guard cannot see the underflow

[xroche]

Seven sites in htswizard.c seed a size_t index with strlen(x) - 1 and then guard the walk-back with (i > 0). SIZE_MAX satisfies that, so the check the author meant as the emptiness test does nothing. They now seed from hts_lastcharoffset(), which #819 adds. On any non-empty string it is the same value, so the only behaviour that moves is the empty case that used to read below the buffer.

Nothing reaches them today. ident_url_absolute() returns -1 on an empty adr and substitutes default-index.html for an empty fil, and that is where the whole invariant lives, since no caller checks fil itself. A producer of an empty fil added later would turn three of the sites live at once.

punycode.c is the exception and the only out-of-bounds write in the set. RFC 3492's sample harness wants both PUNYCODE_COSTELLO_RFC3492_INCLUDE_TEST and ..._MAIN, and nothing in the tree or the build defines either, so it never compiles. Define both and a stdin line whose first byte is NUL leaves strlen() at 0, which segfaults on the read at input[input_length] and would write there next. Checked under ASan against both versions.

Every site being latent, the barrier against the idiom coming back is the source guard rather than a crawl: the self-test picks up the size_t spelling and 01_engine-lastchar.test grows a grep for it.

Based on #819 for the helper. A live crash turned up in the same sweep and went out separately as #829.

Closes #821