Close the coverage gaps a mutation review found, and make the coverage contract enforceable

[creasty]

Follow-up to #65. An adversarial review applied 82 mutations to defaults.go and ran the suite against each: 21 survived, and three tests turned out to be structurally incapable of failing. Test and CI only — defaults.go and setter.go are byte-identical to master.

Three tests that could not fail

• uint8 default:"256" asserting the zero value cannot detect a widened bit size, because uint8(256) is 0. Every integer case passes its width to strconv as a literal, so 14 of the 21 survivors were width and base mutants — int, uint, uintptr, int32, uint16, uint32 and float32 had no boundary test at all. Replaced with the largest and smallest values that must parse, plus past-the-boundary values chosen to truncate to something non-zero (257, 65537, 4294967297, 1e39), and uintptr added to the base subtests.
• TestSet_JSONUnmarshalerSkipsEmptyContainers never evaluated the guard it claimed to pin. It used a type that also implements UnmarshalText, so text won first and the JSON branch was never reached. With a json-only type the two are not symmetric: {} is withheld from a custom UnmarshalJSON, while [] reaches it through encoding/json. That asymmetry is live behavior and is now pinned.
• assert.Zero on the cross-package unexported field could not fail, because that field carried no default tag — nothing would have set it even if reflect were allowed to. It has one now, so removing the CanSet guard fails the test.

Plus one claim the suite cannot make: which of ParseDuration and ParseInt runs first is unobservable, since the two accept disjoint inputs — flip them and every assertion still passes. The test now pins the outcomes and says so, and TestSet_DurationIsTriedBeforeInteger is renamed accordingly.

Shapes that had no test

Shape	Behavior now pinned
[2]int, [1]Struct	reflect.Array has no case: the tag is dropped without an error and elements are not recursed into, unlike a slice's
[]byte	JSON array, base64 string, and bare text being an error
time.Time	UnmarshalText succeeds; an unparsable value is a hard error
non-nil interface{}	not descended into, even holding a *struct with tags
unexported composite fields	the CanSet guard skips a whole subtree, not just a scalar
map[string]*[]T, map[string]*map[string]T	map values pointing at a container are descended into
map element recursion	the parent's tag is not re-applied per element (the slice equivalent was pinned; neither map path was)
*level with a SetDefaults	the one shape whose setter is called only by the pointer branch — so a plausible fix for #67 would silently break it
a second Set	setters fire again: 1→2 for a value field, 2→4 for a pointer field
embedded struct inside a container element	restored from the old suite, which asserted it and the rewrite dropped
[]int default:"{}"	the other decode failure: *json.UnmarshalTypeError, not a syntax error
default:" "	a space is a value, so a string takes it and every container fails to decode it

Coverage is now enforceable

CLAUDE.md says statement coverage stays at 100%, and nothing could fail on a drop: both Codecov statuses are informational: true, and the summary step always exits 0. Delete a test file and CI stayed fully green.

make cover now fails below the threshold, and it profiles the library package alone — under ./... the total read 91.7%, diluted by example/ and internal/fixture, which have no tests, so the number CI printed was not the number the contract is about.

The Codecov upload also loses its matrix.go == '1.27' condition: a literal version silently stops matching the day 1.27 leaves the matrix, and with the token in place there is no reason to single out one leg.

Verified

Each claim above was run, not reasoned:

• 9 of 9 spot-checked mutants now killed — the int/uint/uintptr widths, the uintptr base, float32's bit size, the {}/[] guard, the CanSet guard, the map pointer-element kinds, and the map element tag.
• The gate was tested against a real drop: removing set_error_test.go takes coverage to 95.5% and make cover exits non-zero. (Removing set_map_test.go still passes at 100%, correctly — those statements are redundantly covered.)
• go vet ./..., golangci-lint run, make test, make cover, and the suite under GOTOOLCHAIN=go1.21.13, all clean. 258 test cases, up from 179.

Not in scope

The review also confirmed a crash that cannot be an ordinary test, since the test binary dies with it: a self-referential type whose default materializes an element recurses until the stack overflows. Filed as #71, with the pointer, slice and map spellings.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JjAoN3BBvc6dfT9pog3kMB

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (acc587e) to head (63cd069).

Additional details and impacted files

@@            Coverage Diff            @@
##            master       #72   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files            2         2           
  Lines          111       149   +38     
=========================================
+ Hits           111       149   +38     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.