mod_proxy_balancer: test coverage for #628 (strict balancer-manager numeric parsing)

[notroj]

Test coverage for #628, which replaces atoi()/atof() in the balancer-manager
parameter handling with range-checked parsing. The branch is a new pyhttpd
test that fails on trunk, followed by the two commits from #628 applied as-is,
after which it passes.

Analysis by @claude:

Evaluation. All 14 atoi()/atof() sites in
balancer_process_balancer_worker() are replaced with apr_strtoi64()/strtod()
wrappers rejecting trailing garbage, ERANGE and out-of-range values. Static
helpers, no MMN impact, minimal; 2.4.x has identical code so it backports cleanly.

The semantic fix is more than "hardening": ap_proxy_set_wstatus() treats any
nonzero as set, so atoi("junk") == 0 used to clear a worker flag and "2"
used to set one. Garbage silently mutated state; it is now ignored.

Worth noting on the PR, none blocking:

• strtod("nan") still passes balancer_parse_lbfactor() — NaN compares false
  against both bounds, so (int)NaN is UB. Pre-existing with atof(), but the
  PR claims strictness and leaves this open ("inf" is rejected correctly).
• Boundary nuance: 100.004 was accepted before (int-truncated to 10000) and is
  now rejected (10000.4 > 10000). Harmless; range-checking the truncated value
  would be exactly equivalent.
• Rejections are silent (no log) — pre-existing style.

Proof. test/modules/proxy/test_07_balancer_manager.py drives the manager's
real path (POST + Referer + pinned nonce, state read back from the ?xml=1
view). Case 001 is a positive control, since the handler silently drops
parameters on any nonce/Referer problem.

Before (3 failed, 30 passed):

w_status_D=2 was accepted as 'set': ['Init', 'Dis']
w_status_D=junk cleared the flag: ['Init', 'Ok']
w_lf=1.5junk was applied: {'http://127.0.0.1:5002': (['Init','Ok'], '1.50'),
                           'http://localhost:5002':  (['Init','Ok'], '1.00')}

After both #628 commits cherry-picked unchanged: 33 passed, no regressions.

Two gotchas the test documents: recalc_factors() pins a single member's load
factor to 100, so w_lf is only observable with two members; and a second
member on the same host:port is shared with the first worker, so it must differ
in hostname.

🤖 Generated with Claude Code