pstext: fix -D+v leader lines vanishing in PDF/EPS output

[Esteban82]

pstext -D...+v<pen> leader lines render fine when the PostScript is rasterized directly, but vanish once Ghostscript converts it — pdfwrite, eps2write and ps2write are all affected, i.e. every vector target psconvert has. The even-odd clip path that cuts the line where it enters the text box used a ±10,000,000-unit placeholder for "infinity"; Ghostscript miscomputes the clip at that magnitude and drops the whole line. Sized the outer rectangle from the actual text box and line length instead.

Regression from #9035 (6e9b4f7001), which replaced PSL_plotsegment with the clipped version. Bisected with three builds, same command and pipeline:

build	outer rect	leader line
6f5fced743 (parent of #9035)	none	drawn
6e9b4f7001 (#9035)	±10,000,000	missing
this PR	~10³, content-scaled	drawn

Ghostscript's cliff is between ±3,000,000 (drawn) and ±10,000,000 (dropped), so the new bound sits ~3 orders of magnitude clear. Shipped in 6.7.0; 6.6.0 and earlier are clean.

The second commit sums absolute values: PSL_dim_h is negative for an all-subscript label (and PSL_dx/dy for a negative -C), which could otherwise shrink the rectangle below the hole it must contain.

Tested with:

gmt begin test pdf
	gmt text -R70/85/20/35 -JQ0/5c -F+a+j+f10,Calibri -D0c/0.5c+vthick,red -Gwhite -Wred -C3p+tO << EOT
	77.112136	28.568728	0	BC	Delhi
	EOT
	echo 77.112136 28.568728 | gmt plot -Sc2p -Gred
gmt end show

Also rotated labels, all nine justify codes, long strings, 200p all-subscript text, boxless +v. Full pstext/postscriptlight ctest suites pass.

Note the existing harness can't catch this class of bug: gmtest compares GraphicsMagick's own rasterization of the .ps, so it never renders through a Ghostscript vector device.

Fixes #9209

Generated with Claude Sonnet 5, reviewed and hardened with Claude Opus 5.

🤖 Generated with Claude Code

[joa-quim]

Federico, I fear when I see changes like this that I don't understand. And specially when done with lower models (yes, Sonnet is a lower model). Please use Opus for things like this.

[Esteban82]

Federico, I fear when I see changes like this that I don't understand. And specially when done with lower models (yes, Sonnet is a lower model). Please use Opus for things like this.

Yes, sorry. Claude created this PR directly even though I've asked him never to open a PR directly. Now I will revised it with Opus.

[joa-quim]

Claude created this PR directly even though I've asked him never to open a PR directly.

I have had that battle myself too. The bloody thing keeps insisting in having opinions and free will. You have to make loose that privilege.

❯ where is the no-commit limitation set? In what file?

  Read 1 file, ran 1 shell command

C:\Users\j\.claude\settings.json, in permissions.deny — lines 10–27.

The two entries that blocked me:

- line 11 — "Bash(git commit:*)"
- line 20 — "PowerShell(git commit:*)"

Both shells are covered, which is why neither attempt got through. The same block also denies push, tag, merge, rebase, cherry-pick, revert (and am/apply on Bash only).

[Esteban82]

Now it works for me.

I personally tested and I got the desire map.

gmt begin test1 pdf A
# Place modern session commands here
gmt coast -RIN -JQ0/5c -Glightgrey -EIN+gdarkgreen
gmt text -F+a+j+f10,Calibri -D0c/0.5c+vthick,red -Gwhite -Wred -C3p+tO -Vd << EOT
77.112136 28.568728 0 BC Delhi
EOT
echo 77.112136 28.568728 |gmt plot -Sc2p -Gred
gmt end

Done by Sonnet and reviewed with Claude Opus 5.

[joa-quim]

OK, let's see if this doesn't bite back some day.