I'd say that all the times from 91 degrees onward are off.  Graphing the 
data shows it "jumps" at that point, but is smooth before and after.

Could it be some error related to line 3930 in ext/date/php_date.c

rs = timelib_astro_rise_set_altitude(t, longitude, latitude, altitude, 
altitude > -1 ? 1 : 0, &h_rise, &h_set, &rise, &set, &transit);

Altitude is 90-zenith, so when you get to 91, the code would set that 
param from 0 to 1 ... which I think means it calculates the twilight 
time instead of sunrise/set.  I would say this param should always be 1, 
if you are calculating sunrise/set (see note on line 187 of 
ext/date/lib/astro.c).

TYPO IN LAST COMMENT - SHOULD READ:

Altitude is 90-zenith, so when you are below 91, the code sets that 
param to 0 ... which I think means it calculates the twilight 
time instead of sunrise/set.  I would say this param should always be 1 
if you are calculating sunrise/set (see note on line 187 of 
ext/date/lib/astro.c).

i think the bug is being caused by this 
altitude > -1 ? 1:0 ,line 3930 in php_date.c. 
becoz of this upper_limb correction is being applied here....
if (upper_limb) {altit -= sradius;} line 253, astro.c 

the series gets advanced because of the altit-=sradius;. 
and hence the anamoly. i dont knw if its a bug. don't know about the upper limb correction and why is it being selectively beyond -1 applied. but the bug can be removed if we remove the uppr_limb if code. or add it without an if. and if the correction is correctly applied, then i guess thats the way it is.not a bug.

Hardcoding that upper_limit to correction flag to either 0 or 1 does remove the quirk in the series.  So, is it correct that for the sunrise function we should just peg it at 1?

This is from the original sun.c file
Written as DAYLEN.C, 1989-08-16
Modified to SUNRISET.C, 1992-12-01
(c) Paul Schlyter, 1989, 1992
Released to the public domain by Paul Schlyter, December 1992
Portions Modified to SUNDOWN.NLM by Cliff Haas 98-05-22

the comment states
/*         upper_limb: non-zero -> upper limb, zero -> center       */
/*               Set to non-zero (e.g. 1) when computing rise/set   */
/*               times, and to zero when computing start/end of     */
/*               twilight.                                          */

Therefore, if we are calculating sunrise-sunset time the value of upper_limb will be set to permanently non zero.
If twilight time is needed then it has to be 0.


(PS - just for clearing what this upper limb is. Sunrise happens when the sun's centre is visible. But for the twilight to occur, the sun's top should come over the horizon. that means,centre angle - sradius:)

The comments from the sun.c file are also in PHP's ext/date/lib/astro.c 
(around line 187).

From my reading of things, this upper-limb value should be pegged at 1 
in php_date.c

Which, of course, means that most systems have been generating 
sunrise/set times that are off by about 26 seconds.  Think of all the 
angry emails you're gonna get from The Weather Network!

Fixed in SVN