fix(light): classic zi light loads with full tracking, like zi load

[ss-o]

Environment

• Zi next at 3fa8433 (GitHub tarball), Zsh 5.9.2, Linux x86_64, zsh -f.

Minimal reproduction

A local plugin directory demo holding demo.plugin.zsh:

demo_fn() { :; }
bindkey "^X^D" demo_fn

Load it once in each of three fresh zsh -f shells, each sourcing zi.zsh, then print the plugin's state records from ZI:

zi light /path/to/demo            # shell 1
zi light-mode for /path/to/demo   # shell 2
zi load /path/to/demo             # shell 3
for k in ${(k)ZI}; do
  [[ $k == (STATES__*|FUNCTIONS_BEFORE__*|BINDKEYS__*) ]] && print -r -- "$k=${ZI[$k][1,40]}"
done

Actual behavior

Command	STATES__%.../demo	FUNCTIONS_BEFORE__	BINDKEYS__
zi light /path/to/demo	2	recorded	^X^D demo_fn recorded
zi light-mode for /path/to/demo	1	empty	empty
zi load /path/to/demo	2	recorded	^X^D demo_fn recorded

Classic zi light loads the plugin with full tracking, exactly like zi load. The classic-syntax branch in zi.zsh sets ICE[light-mode]= (line 3030 on next) and then runs ICE=() ZI_ICE=() (line 3035) after copying ICE into ZI_ICES, so the light flag does not reach the load.

Expected behavior

zi light loads without tracking, as README.md:75 ("light loads a plugin without tracking and reporting overhead") and docs/man/zi.1:562 ("Light plugin load, without reporting/tracking") say: the same records as zi light-mode for (STATES__ 1, no function or bindkey records). zi light -b should keep tracking bindkey calls only.

Found while reviewing the unload work for #113 and #582; it is not part of those issues.

[ss-o]

Project 28 triage. Values set on this new issue:

• Status Todo: reproduced on next at 3fa8433 (table above), and the documented behaviour (README.md:75, docs/man/zi.1:562) defines the fix.
• Item Type Bug, Impact Medium: every classic zi light user pays the tracking cost the command exists to avoid, and unload records differ from light-mode for.
• Effort S: carry the light flag through the classic-syntax branch (zi.zsh lines 3030-3035) plus a test comparing the three load forms.
• Priority P2 - Medium, Workstream Human delivery: the same values as the sibling unload issues fix(unload): repeated-load takeover misses linked keymaps and duplicates keymap records #583 and refactor(unload): chain-aware restore for per-load ownership records #585.